Beyond the Slide

Autopsy #002: Two Bodies in the Molecular Window

Dr. Luis Cano — Fri, 30 Oct 2026 10:58:36 GMT

Like every end of the month, I have dues to pay. My duties demand it, and so does my own passion for the specialty: the autopsy room is waiting. I arrive early, as always. On my phone there is a playlist of old Japanese music that I have put together with care on Spotify. My over-ear headphones shut me off from everything and everyone. I clear my mind, leave my street clothes behind and put on my scrubs. Then I unplug the headphones, pick up the phone and let the music play out loud. I sling a belt bag across my back, and there I go... the room in front of me... what will I find this time, what story will today’s body tell me... but... I think today will be different. There isn’t one. There are two bodies, and they were found in the same room.

The Orphan Question

Dr. Luis Cano — Fri, 23 Oct 2026 09:58:34 GMT

Last week I joined a panel at Pathivision, the topic was integrated diagnostics: how to organize so much information in a way that creates value for the patient, for the physician, for science and for industry.

The discussion was interesting, until someone from pharma asked a question at the end that caught my attention: how can integrated diagnostics improve the effectiveness of the industry, without ending up answering complex problems with even more complex solutions?

I took the question and proposed my three-question framework for guiding integrated diagnostics development in pharma. I waited for a rebuttal, or at least a comment. I heard “interesting,” and the conversation moved on.

That silence stayed with me. Had I been that clear? (Which I honestly doubt: I know myself.) Or was the answer already, at least in part, known to everyone? And if so, two more uncomfortable questions followed. What could I add that was new? And above all: if the answer is known, even partially, why is it applied so rarely?

The comfortable hypothesis

Let me start with the three questions, as I presented them.

The first: why are we doing this, and what are we looking for? This is the question of purpose and need. It sounds obvious, but in practice it gets answered with sentences that are not answers: “because the technology exists,” “because a competitor is doing it,” “because there is budget this year,” “because the biomarker is biologically interesting.” None of them says what we expect to gain, or for whom.

The second: if we do this in a controlled environment, will it survive the real world? This is the question of whether the result holds up under the conditions in which it will actually be used. Every integrated diagnostic is born under ideal conditions: one central lab, one scanner, one staining protocol, a handful of pathologists trained to the same standard. The real world is dozens of labs, different platforms, irregular fixation, small biopsies, and general pathologists signing out eighty cases a day.

The third: what decision does the result change, in the real world? Not what it detects, nor how accurate it is, but what the treating physician does differently on the day that result is in their hands, and what benefit we expect from that change. Changing a decision is not enough; it has to change it in someone’s favor. If the answer is “nothing,” the integrated diagnostic is, more than anything, an exercise in unnecessary sophistication.

The most comfortable explanation for why diagnostic integration fails so often is that nobody asks these questions in time. (I wish that were true: I could sell more of my framework and get more work, but reality, as always, is more complicated.) Programs start out of technological enthusiasm, a recommendation, a messianic idea, or an interest nobody ever fully understands. Then the trial is designed around the drug, and by the time someone asks what the test is for outside the trial, it is too late to change much.

If that explanation were right, the solution would be simple: teach the questions. Put them on a slide, present them at congresses, add them to project approval templates.

But as my grandmother used to say, when something sounds too simple, be suspicious. Either nobody understood what is going on, or everyone already has the answer. So I decided to be suspicious.

The questions are 35 years old

I wanted to see whether anyone had formalized something similar. I did not have to look far.

In 1991, Dennis Fryback and John Thornbury published a hierarchical model for evaluating the efficacy of diagnostic imaging, in six levels: the test works technically, it is accurate, it changes the physician’s reasoning, it changes treatment, it changes patient outcomes, and it is worth what it costs society. It was born in radiology, but it became a reference for evaluating diagnostic technologies in general.

Then came the CDC’s ACCE model for genetic tests (analytic validity, clinical validity, clinical utility, and ethical, legal and social implications) and, in 2009, the EGAPP initiative, which built on that foundation a formal method for deciding which genomic applications deserved to reach practice. In different words, they all say something similar: accuracy is necessary, but not sufficient.

My questions do not map neatly onto these frameworks, and I would rather say so. The first does not appear in the hierarchy because it comes before it: it defines what we are evaluating for. The second resembles external validity, which is not the same as accuracy measured under controlled conditions. The third comes close to clinical utility, the levels at which a test changes treatment and, later, outcomes. I did not invent the substance. I reframed it for a context, integrated diagnostics in pharma, that these frameworks did not have in mind.

Maybe that explains the silence in the room. Or maybe not; I have no way of knowing. But I do know this: these questions have existed, been taught and been cited for decades (no more work for me then… or is there?). The failure of diagnostic integration cannot be explained by a lack of knowledge alone. My suspicion is that the problem lies in how the questions are divided up: who asks each one, and when.

The orphan question

In many drug–diagnostic co-development programs, each question tends to land in different hands, at different times.

The first is usually answered by translational medicine or the biomarker lead, early on. Their focus is biology: whether the target makes sense, whether the biomarker separates responders from non-responders. That is a legitimate answer, but it explains why the biomarker is interesting, not what the test will be for.

The second usually sits with the diagnostics team and the partner manufacturing the test, during clinical development. That is when the assay lock arrives: the test is fixed to select patients for the pivotal trial, with that central lab and those pathologists. Whether it will survive in a community hospital tends to be left for later.

The third is the one most often left without an explicit owner. In theory it belongs to everyone. In practice, it often reaches medical affairs or market access after the trial has ended, when the variables a payer or a hospital would need can no longer be collected. This is not negligence. It is that, at the moment when it can still be answered, it is rarely on anyone’s agenda.

Writing this, I understood something I had actually lived through, and which was the origin of the framework. When I built it, I knew each phase was led by a different team, and I thought the three questions could bring them together. What I never specified (my mistake) was where and with whom to use them. And that was the point. It is not about each team answering its own question and passing the baton to the next. All three questions have to be asked at the same time, by all the teams or their representatives, at the moment the race for the biomarker begins.

And that is the problem: when these questions are run as a relay race, the third arrives at the end of the course, when there is little left to change. Not because anyone thinks it unimportant, but because the relay leaves it for the final leg.

The Orphan Question: the question left without an owner when all three are run as a relay race, instead of being asked together at the start.

PD-L1 in lung cancer shows the cost of that relay in harmonization and implementation. Each immunotherapy was developed with its own antibody, its own platform and its own cut-off, and each trial answered its second question well within its own controlled world. What remained open was the question of the real laboratory: what does a pathology service do when it must report results for several drugs and cannot maintain five different assays?

That question was eventually taken up by the Blueprint project, coordinated by the International Association for the Study of Lung Cancer (IASLC), with academia and pharmaceutical and diagnostic companies at the same table, after the approvals. They found that three assays were reasonably comparable on tumor cells, and that scoring on immune cells was far less reproducible. In other words, the questions did end up being discussed together, but only once the problem was already in the labs.

HER2 shows the same dynamic in reproducibility. Antibody–drug conjugates moved the treatment decision toward HER2-low and then HER2-ultralow, tumors with faint membrane staining that used to be reported as zero: a threshold immunohistochemistry scales were never designed for, and whose reading fell to the pathologist in each hospital. In a multicenter study, agreement between pathologists in separating HER2-null from HER2-ultralow was poor (Fleiss kappa 0.230). Other work shows that immunohistochemistry remains reproducible overall (kappa 0.67–0.77) and that the problem concentrates in borderline cases. But the borderline is exactly where it is now decided who receives the drug.

Archaeology, not design

When the third question arrives at the end of the relay, it often ends up being answered backwards. You can see it even in cases the industry rightly presents as successes.

Oncotype DX is probably the most cited example of a genomic signature that changed practice. It did not lack evidence: its prognostic value was validated in 2004 on archival samples from a historical trial, and in 2006 a retrospective analysis of another trial suggested which patients benefited from chemotherapy. The test reached the market in 2004. But when the EGAPP group reviewed the evidence, it found no direct proof that guiding treatment with the test improved outcomes. Prospective, randomized evidence arrived in 2018, with TAILORx. Fourteen years between launch and the prospective answer to the third question.

TROP2 illustrates the value of looking backwards, and its limits. The phase III TROPION-Lung01 trial enrolled lung cancer patients without selecting them by TROP2. Afterwards, in an exploratory, retrospective analysis, a computational pathology tool, quantitative continuous scoring (QCS), was applied to the samples. It measures how much TROP2 sits in the membrane versus the cytoplasm. Patients classified as positive by this method had a clear benefit over docetaxel; negatives did not.

It is remarkable work, and I say that as someone who believes in computational pathology. But the third question was not designed: it was found in the data of a completed trial. Its value as a selection tool still has to be confirmed prospectively.

In oncology genomics, a type of study called the decision impact study has also grown: studies that measure whether a test changes the treatment recommendation. A 2023 bibliometric analysis concluded that this evidence is mostly conceived and produced by industry. That does not invalidate it, but it calls for careful reading: measuring a change in recommendation is not the same as demonstrating a benefit.

The villain of this story is not a company or a person. It is a development architecture built around one milestone: drug approval. Whatever helps reach that milestone has an owner, a budget and a timeline. What happens afterwards, in the hospital, tends to be left for later. And “later” is precisely when the third question gets answered worst.

When someone signs

There is one place in pathology where diagnostic synthesis does have an explicit owner.

In 2016, the WHO classification of central nervous system tumors introduced the layered integrated diagnosis: a top line built from histology, grade and molecular information. The 2021 edition deepened the change, and between editions the cIMPACT-NOW consortium kept updating the criteria so that practice would not have to wait five years.

The consequences are concrete. A diffuse IDH-wildtype astrocytoma that looks low grade under the microscope, but carries a TERT promoter mutation, EGFR amplification, or the combination of chromosome 7 gain and chromosome 10 loss, is diagnosed as a grade 4 glioblastoma. Morphology and biology can contradict each other, and someone has to resolve that contradiction. That someone has a name and signs the report: the neuropathologist. Integration came at a cost in turnaround time, training and friction between services, but the final synthesis has an owner.

Molecular tumor boards offer the contrast. They bring together oncologists, pathologists, geneticists and bioinformaticians to translate genomic findings into recommendations. A 2025 meta-analysis pooled 34 studies with more than 12,000 patients, and a 2026 meta-analysis concluded that randomized evidence supports a clinical benefit, though limited for overall survival. But the board recommends; it does not decide. The decision stays with the treating oncologist, who has to secure the drug, the trial or the authorization. Between recommendation and prescription lies a gap that often has no one assigned to it.

Diagnostic integration works better when someone signs it than when everyone recommends it.

I leave open one question for which I found no sufficient evidence: when a diagnosis integrates histology, molecular data, imaging and clinical information, does the real-world variability of each layer add up, multiply or cancel out? My intuition as a pathologist is that it accumulates, but I would rather leave it as a question than sell it as a certainty.

The pool

At this point it would be easy to close with a promise: ask the three questions together, at the start, and diagnostic integration will work. It would be a neat ending, but also the most simplistic and absurd way to end.

BFAST shows why. It was the first prospective randomized trial designed to evaluate blood-based tumor mutational burden as a predictive biomarker for immunotherapy in lung cancer. The biomarker, the threshold and the decision were defined by design, before the first patient was treated. And the biomarker did not deliver: in patients with high mutational burden, the trial missed its primary endpoint (HR 0.77; p = 0.053).

This does not weaken the argument; it puts it in its place. Asking the questions well lets you test a hypothesis cleanly, not guarantee that it is right. What it does is increase the chance of spotting certain problems earlier and deciding with better evidence, with less money spent and fewer patients selected with a test that is hard to reproduce.

It won’t keep your feet dry. It will keep you from diving in head first without knowing whether the pool is full or not.

That is why I distrust any framework, mine included, that presents itself as a solution. The three questions improve your odds, standardize the conversation and make visible what would otherwise be discovered late. But improving your numbers is not the same as guaranteeing them.

The questions, returned

Back to the room at Pathivision. I still do not know whether people nodded because they already knew, or because I did not explain myself well. But I now know what my answer was missing: I never said when, or with whom, to ask the questions.

The questions do not change. A useful question is one you remember in the hallway, and these are remembered because they are simple:

What are we doing this for?
Will it survive the real world?
What decision does it change, and who benefits?

What changes is how they are used. Two adjustments, and both fit on a napkin.

First: ask them together, at the start. Translational medicine, diagnostics, clinical development and market access, in the same conversation, before the race for the biomarker begins. Not as a relay.

Second: put a name on it. The conversation belongs to everyone; the follow-through belongs to one person, with real authority to bring the teams together, ask for evidence and carry open questions through to a decision. Without that, the name is just another box on a template.

You do not need a new framework or a forty-field template. You need one conversation at the start and someone to sustain it. It is a small thing, and that is exactly why it is so easy to skip.

So I will leave the question I should have asked in that room:

Who in your organization owns question three?

Can AI Make Us More Human?

Dr. Luis Cano — Fri, 09 Oct 2026 09:59:46 GMT

A while ago I went with a friend to a medical appointment. On the way, almost in passing, he told me what was wrong.

For more than eight months he had been living with chronic pain in his abdominal wall, under the right ribs. First they said it was his gallbladder, they removed it. The pain stayed. Then came the paracetamol prescriptions, one after another. They told him it was stress. They let him understand he was imagining it.

I asked how long it had been going on. I asked whether he had seen another specialist. His answer carried a resignation I still think about. Everyone told him the same thing. No one had examined him. “They don’t even look at me,” he said. “They give me a prescription or send me for tests, and that’s it.”

I am a physician. I don’t practice in the country where I live, but listening requires no license, and physical examination is not a sin. So I asked whether he wanted me to examine him.

The pain responded to a specific sign: a positive Carnett’s sign, the one that points to the abdominal wall rather than the organs beneath it. I recognized it because I had seen it years earlier, during my rural service year in Peru, in places where clinical examination is your only friend, and your best one. The picture pointed toward anterior cutaneous nerve entrapment syndrome, ACNES. I found a couple of papers and sent them to him. He read them, recognized himself, and took them to his doctor, who set them aside.

I am not writing this to point at that doctor. I am nobody’s judge. Where I live, things are done differently, within a different system. But one variable repeats wherever I look: the time the system leaves for each patient. That is what this essay is about.

A few days ago I ran into him again. He was still looking for someone to treat him, now abroad. Before we said goodbye, he told me something else: “Of all the people I saw, you were the only one who really listened to me. The only one who told me what it might be, and who cared. Thank you.”

I did nothing extraordinary. I did what my profession taught me: listen, ask, think, and when you don’t know the answer, go look for it. Connect. Understand. That is medicine too. I don’t know whether we have lost it, or whether we are entering a spiral where doing more is the norm, even when doing better doesn’t come with it.

Maybe it was coincidence, but that same week, scrolling through my feed, I came across a headline: AI will make medicine more human.

I couldn’t help noticing the irony. How is software going to teach you to be more human? If the software is the software and the human is you, isn’t the easiest way to be more human simply to be human?

I decided not to stop at the headline, digging further, I came across a book about exactly this, written by a physician widely considered one of the most influential minds in medicine today: Eric Topol. It caught my attention, and I read it cover to cover, It is called Deep Medicine, and its subtitle reads almost like a manifesto: AI can make healthcare human again. The question came back to me: how can a program help me humanize the most human profession there is?

I confess that was my bias going in. But as I read, I realized the headline didn’t do the argument justice. The argument is elegant, and largely right in its diagnosis.

Somewhere in the last few decades, the physician became a data-entry clerk. More time facing the keyboard than the person in the room. Exactly what my friend had felt. If machines take over that burden, they give us back what Topol calls the gift of time. And returned time goes back to its rightful owner: the patient, looking again, listening again, touching again, connecting again.

As a good pessimist, I think it is a hard promise to keep, not an impossible one. Anyone who has worked in a hospital knows that dehumanization is not a metaphor. It is a specific workday, with a specific number of cases, an inbox that never empties, and notes finished at night. Any tool that eases that deserves to be taken seriously, very seriously.

But medicine has two fronts, the human one, which is where it began, and the functional one, which is how it is built today. A few months ago I argued here that the real market for clinical AI wasn’t diagnosis but productivity: everything that steals hours from a specialist without needing a regulator-approved algorithm. I still believe that. But that analysis looked only at the second front, medicine as an engineering process, with inputs, outputs and bottlenecks. This essay looks at the other one.

Because engineering has a limit that medicine knows well. If I have a task list and a fixed number of hours, the productive thing is to do each task as efficiently as possible. But you cannot be efficient with pain, You cannot tell a patient they have three seconds to cry, ten to explain what is happening, and five to go through their list of symptoms.

Topol’s promise answers that tension with a single variable: time. Which gives it one great virtue. It can be tested. Time can be measured. And over the last two years, for the first time, we have started to get data.

Does it give us time back?

If any tool should deliver the gift, it is the ambient scribe. It listens to the visit, understands the conversation, and drafts the note while the physician looks at the patient. It diagnoses nothing, It interprets no image, It does exactly what the promise needs: it takes the keyboard out of the room.

The largest study so far followed more than 8,500 ambulatory clinicians across five US academic health systems between 2023 and 2025. About 1,800 adopted the scribe. The study used passive EHR logs, not surveys. It measured what physicians did, not what they felt.

The result is real, and it is small.

For every 8 hours of scheduled patient time, adopters spent 16 fewer minutes on documentation and 13.4 fewer minutes in the EHR: roughly a 10% and a 3% relative reduction.

A randomized trial at UCLA with 238 physicians compared two commercial scribes. One reduced time per note by 9.5%; the other showed no significant reduction.

Sixteen minutes across a clinical day, not trivial, maybe two conversations that didn’t have to be cut short. But not the hours the promise implied. The gift exists. It just fits in a pocket.

So far, the promise survives, a quarter of an hour returned to the patient each day is still humanity recovered.

As long as it goes back to the patient.

Where does that time go?

Reading that study, something else caught my eye. This is where the promise starts to crack. The researchers didn’t only record how much time was saved. They also recorded what happened to the schedule.

Adopters saw, on average, half a patient more per week. A modest, statistically significant increase, which the authors translated into additional revenue per clinician. Meanwhile, after-hours EHR work, what physicians call pajama time, didn’t change. The researchers suggest the time may have gone to patient messages or to reviewing the AI’s own drafts. The study didn’t measure it.

Scribe adoption was associated with 0.49 additional visits per week (a 1.7% increase in weekly volume) and about $167 per month in additional visit revenue per clinician.

After-hours EHR time did not change significantly.

Let’s be precise about what this means, nobody sat in an office and decided that the time returned to physicians would become volume, nobody had to. When the unit a system uses to measure itself is the visit, free time flows toward the visit the way water flows downhill. It isn’t a conspiracy, It is institutional gravity.

And it isn’t the first time…

In 1983, historian Ruth Schwartz Cowan published More Work for Mother. Her question was simple. If the twentieth century filled homes with washing machines, vacuum cleaners, electric stoves and running water, all designed to save labor, why weren’t full-time homemakers working less? In 1974, sociologist Joann Vanek had reached a puzzling conclusion: women who didn’t work outside the home spent as much time on household work as they had fifty years earlier. Decades later, researchers went back to the original diaries and refined the picture. Time spent cooking and cleaning did fall. But much of it moved to childcare and shopping. Technology saved time on some tasks, and that time went to others.

Cowan’s explanation wasn’t that the machines failed, They worked. Each task took less effort, what changed was everything else. Technology made the help disappear: the servants, the share once carried by other family members. And it raised the standards: cleaner clothes, more elaborate meals, more hours of childcare, new tasks that hadn’t existed before. The machine saved the effort of each task, the system kept the hours.

The analogy has limits, a home is not a hospital, and Cowan proves nothing about AI. But she describes a mechanism medicine already knows. The electronic health record also promised to save time, a direct-observation study found physicians spending nearly two hours on desk work for every hour of direct patient care: 27% of the clinic day face to face with patients, 49% on the EHR and desk work. It doesn’t prove causation. It does show how the last promise ended.

This pattern deserves a name: The Allocation Gap.

Any time-saving technology goes through four stages. First, a specific task gets faster. Second, total workload falls, or doesn’t, once you account for reviewing, correcting and learning the tool. Third, the saved time gets allocated to something: more patients, more complex cases, a longer conversation, teaching, rest. Fourth, all of that produces outcomes: quality, safety, wellbeing, patient experience.

The industry measures stage one carefully, It sells stage four, almost nobody measures stage three. And what goes unmeasured isn’t left free. It gets allocated, by default, to whatever is measured. In a system that counts visits, unclaimed time becomes one more visit.

Here the original question starts to break. Asking whether AI can make us more human assumed time was the obstacle. But returned time doesn’t reach the patient on its own. Maybe we were asking the wrong thing. Maybe there is a prior question: what are we optimizing?

What are we optimizing?

The two fronts I mentioned at the start are, in practice, two parallel realities. And almost all the technology we have built optimizes only one of them.

The first is the procedural reality, It has a clear center, diagnosis, and a protagonist, the physician, who must produce correct diagnoses within a given time. It has metrics: accuracy, turnaround, cases per day, cost per case. Procedurally, diagnosis is what matters most, and this reality deserves all the engineering we can give it.

The second is the human reality. It begins before the diagnosis and continues long after. Its protagonist is the patient, who bears the consequences of whether the diagnosis is right, and who then has to live with it. This reality has its decisive moments too. The physician who listens or doesn’t. Who looks or doesn’t. Who understands the pain, who offers hope or, when there is none to offer, at least doesn’t walk away.

But this reality has no metrics. No indicator tracks the minutes a physician spent looking the patient in the eye. No dashboard records whether someone was present when the news was bad.

Pathology, where AI has gone furthest in diagnosis, makes the asymmetry unusually clear. ARTICULATE PRO is one of the most serious real-world evaluations of pathology AI to date: prospective, across three NHS centers, with prostate biopsies reported by uropathologists in routine practice. Its results are solid and deserve respect.

Of 1,613 cases, 1,049 were reported with AI assistance. In the second-read phase, AI led to a change in diagnosis or Grade Group in 21 of 386 cases (5.4%), 5 of them (1.3%) with potential impact on clinical management.

Cases requiring immunohistochemistry fell at all three sites. Overall turnaround dropped by 30.1 hours at one site; at the other two, the difference was not statistically significant.

These are real benefits for real patients: fewer extra tests, faster answers, some diagnoses corrected. But the most revealing part of the study isn’t in its tables. It’s in its discussion.

When the authors imagine what the saved time means at national scale, they translate it entirely into system capacity: days of waiting returned to NHS patients, national cancer diagnosis targets, laboratory capacity freed for other work.

These are legitimate goals, nobody in their right mind objects to a man with suspected cancer waiting one day less. But one thing never appears: the pathologist’s time. The study didn’t measure whether the saved hours went to a complex case that deserved more attention, to a tumor board, to training a resident, or simply to the next biopsy in the queue.

I am not criticizing the authors, They designed an excellent study for the questions they asked. It is the grammar of the system: in medicine, saved time is born already allocated.

And the patient does appear, in a very specific form: as days of waiting. As a queue. Present in every metric as a number, and in none as a person. The absent presence of this whole conversation.

We optimize the procedural reality with extraordinary sophistication. The other still has no instruments. Which forces a harder question. Has technology at least solved the problem that matters most on the procedural side: being wrong less often?

Subscribe now

What technology never solved

Between 1959 and 2000, a German university hospital did something very few hospitals do. Decade after decade, it compared clinical diagnoses against autopsy findings. Five medical eras. Over that span came ultrasound, CT and MRI. Diagnostic medicine was transformed.

The misdiagnosis rate didn’t move. In 1999–2000 it was still 11%, as in the decades before. The new imaging techniques did not reduce it, meanwhile, history-taking and physical examination still led to the correct diagnosis in roughly three out of four cases.

This needs an honest reading. It is one hospital. And the share of deceased patients who underwent autopsy fell from 88% to 20% over the period, so later autopsies likely skewed toward harder cases. A systematic review of 53 autopsy series found, after adjusting for these factors, that errors did decline over the decades. They didn’t disappear.

Adjusting for case mix, country and autopsy rate, major diagnostic errors declined 19.4% per decade, and errors likely to have affected patient outcome declined 33.4% per decade.

Even so, the authors estimated that modern US institutions still see major error rates of 8.4% to 24.4%, and outcome-affecting errors of 4.1% to 6.7%.

A more recent national estimate puts at about 795,000 the number of people in the US who suffer serious harm (permanent disability or death) from diagnostic error each year.

The honest reading isn’t that technology failed. It reduced error without eliminating it. And what remains lives in a place machines struggle to reach.

Because medicine, at its core, rests on a very old hypothesis: listen, understand, question. Not questioning to humiliate the patient, or to force what they say into our mental frame, but to free that frame from its own doubts so we can see clearly. The foundation of medicine isn’t lab tests. It is knowing how to navigate a patient’s fear and pain until you find the diagnosis that fits.

The diagnostic error literature has technical names for what happens when that process fails: anchoring, premature closure. The physician finds an explanation early and stops looking. They fit the patient into their frame instead of letting the patient challenge it. An analysis of a hundred diagnostic errors in internal medicine found premature closure was the single most common cause, and that system factors were also involved in two out of three cases. No MRI fixes that. One more question fixes it, asked in time, by someone who was still listening.

One detail in ARTICULATE PRO shows how delicate this territory is. In one case, the pathologist had seen a small suspicious focus, leaned toward calling it benign, and would normally have ordered an extra stain to confirm. But knowing the AI would review the case afterward changed their decision threshold. The authors say it plainly: knowing AI will be available changes the dynamic between pathologist and case. The tool doesn’t just answer questions. It changes how we look.

That can be good or bad. A tool can free our frame from doubt, or it can replace it. It can help us see more clearly, or become the anchor. And the question turns again. If the core of diagnosis was always human, if listening is still the most powerful diagnostic tool we have, then the question isn’t whether AI makes us more human. It is what, exactly, makes us human in a consultation.

What actually makes care human?

The most interesting answer comes, paradoxically, from a trial of AI scribes.

At the University of Wisconsin, a randomized trial followed 66 clinicians for 24 weeks as they gained staggered access to an ambient scribe. It was funded by the health system itself, not the vendor, and it measured something few studies measure: professional wellbeing, with validated instruments.

Work exhaustion fell by 0.44 points (95% CI −0.62 to −0.25) on the Stanford scale. Professional fulfillment, the other primary outcome, did not improve significantly.

In qualitative interviews, patients valued the reduced interaction with the computer and described communication as clearer and more open.

There are two lessons here. First: patients experienced a more human visit without the visit getting longer. The minutes gained were modest, but the physician’s eyes left the screen. What humanized the encounter wasn’t time. It was attention: presence within the same minutes. Topol’s promise bet on the quantity of time. The data suggest what matters is its quality.

Second: less exhausted is not the same as more fulfilled. The tool lifted a weight. It didn’t add meaning. Which makes sense. The meaning of the profession was never in the keyboard. It was in what the keyboard kept us from doing.

We shouldn’t overstate what we know about the clinical value of that presence. A 2011 study found that patients of more empathic physicians had better diabetes control. An attempt to replicate it in a larger cohort found no such association. A meta-analysis of randomized trials found that improving the patient–clinician relationship has a small but real effect on health outcomes. And one telling detail: that analysis excluded trials in which visit length differed between groups. The effect didn’t come from more time. The evidence doesn’t let us sell empathy as a biomarker.

Maybe it shouldn’t have to. Treatment was never just taking medicine. It is also: I hear you, I understand your pain, I am with you, I give you hope. And when I can’t, because the disease is terminal or there is nothing left to do, I stay with you in your suffering. That part doesn’t need to lower HbA1c to count as medicine. It is the part of medicine patients remember.

AI can remove obstacles: the screen, the keyboard, the midnight note. What it can’t do is give meaning to the space it frees. So the question changes once more. It isn’t about AI making us human. It’s about AI letting us be. And whether we will take the opportunity.

Will we remember?

There is an idea in Deep Medicine I want to put on the table, because Topol himself raised it. The efficiency AI brings cuts both ways. It can be used to protect the relationship between physician and patient, or to squeeze physicians even harder. The gift of time was never automatic. It is a decision.

And here I understand my initial discomfort better. By putting time at the center, the promise turns a human problem into a technical one. If humanity is missing, it’s because time is missing; if time is missing, the fix is a tool. But time was never just a variable. It was always a choice about how to spend it.

That choice isn’t made by algorithms. It is made by scheduling templates, payment models, the indicators a department is judged on, the contracts that define what is expected of each clinician once a tool is deployed. In other words, by us: medical directors, department heads, procurement leads, clinicians themselves. And we are making it now, with every deployment, almost always without realizing it is a choice.

Closing the Allocation Gap doesn’t require new technology. It requires a different kind of measurement: not for what we want to standardize, but for what the patient can feel and experience. Not to put a price on it, but to recalibrate its value. Something as simple as asking patients, on their way out, whether they felt heard, and treating that answer as seriously as the documentation minutes saved.

It also requires protecting time explicitly and deliberately, in writing if possible, so the next appointment doesn’t absorb it. And it requires asking, before any contract is signed, where the saved time will go, and checking afterward where it went.

I started this essay thinking the question was absurd. How could software make us more human? Writing it, I realized I was the one seeing only part of the problem. The premise wasn’t wrong because a program claims to teach me what, as a human, I already know. It was incomplete. What technology can do is give us back a few minutes. What it can’t do is remind us what they were for.

So the question that remains isn’t technological. It’s the old one, except now we have no excuse to avoid it. When the machine gives us our time back, will we turn it into one more visit, or one more conversation?

I think about my friend. Nobody denied him modern medicine: he had tests, surgery, prescriptions. What he lacked was someone who would listen to him and examine him. Eight months later, his words weren’t about a diagnosis. They were: you were the only one who listened.

Those words stayed with me. Not for what they say about me, but for what they say about us: that listening, the first thing we were taught, has become the exception. They made me feel like a doctor again. They made me feel human. And they left me with a question that is no longer about algorithms, one I ask myself before anyone else:

Will we be able to remember that we were human?

Autopsy #001 : Oral semaglutide in early Alzheimer's disease

Dr. Luis Cano — Fri, 02 Oct 2026 09:58:46 GMT

Pathology is lonely work that few people want to do. A friend told me that once. After more than ten years at it, I can say he was wrong. What I love about pathology is the urge to know, and above all to find the cause hiding behind what we see on a glass slide. And then there are autopsies. Forensic pathology doesn’t exist under that name in my country. It lives inside legal medicine, where the pathologist doesn’t practice as such. But it’s the subspecialty that has always pulled at me: the method, the way it’s done, and the fact that in the end, the tissue has the last word.

Starting today, once a month, I’m taking that habit somewhere else. No patients, no criminal cases: I’ll be performing autopsies on the pharmaceutical and biotech industry. We’ve all watched the symptoms. It’s time for a final diagnosis.

My first patient lies on the steel table under a sheet: semaglutide, the molecule behind Ozempic and Wegovy, in its 14 mg daily oral form. It died on November 24, 2025, in twin trials, evoke and evoke+, after 104 weeks of treatment in 3,808 people with early Alzheimer’s disease.

As far as anyone knows, there was no visible trauma. No toxicity, no scandal, no manipulated data. The trial did exactly what it promised: it measured. And what it measured was nothing.

That’s the first oddity of the case. Clean bodies are the hardest to read, because the cause of death isn’t in the wound. It’s in the question that brought the body here. I proceed.

The FDA’s Framework Measures How Autonomous the AI Is, Not Whether the Doctor Is Still Thinking.

Dr. Luis Cano — Fri, 25 Sep 2026 06:24:00 GMT

A friend sent me a working paper a few days ago that genuinely grabbed my attention. It’s from the FDA, and its purpose is ... to collect ideas and opinions to help shape what will become the new way of integrating AI into digital pathology (and into medicine as a whole, to be more precise). I obviously couldn’t let it pass, so I sat down to read it as if I had nothing else to do. And what I found, I don’t know how to classify. I don’t even know where to start ... there’s so much, and the document is so dense, that somewhere between the ocean of unanswered questions and the fog of concepts clouding their own meaning, I think I arrived at a realization: AI in medicine, the way it looks right now ... is playing a different game entirely.

The medicine I know works like this: someone walks in, tells you what’s wrong ... you interpret it, you start building your hypotheses, and then come the tests to confirm whether you were right, whether you know what your patient has or you don’t and need to keep looking. So you always keep your books close, plus a colleague, plus a notebook for notes. Not because you don’t want to see the patient, because memory is fragile. That process is, at its core, a human act. Not because it can’t be replaced by an AI agent, but because the human inside that process is the one who’s responsible ... and that’s when I saw it. That’s also where this question was born.

The AI we want to regulate and adopt falls into this frame .... Interpretation ... hypothesis generation ... ordering tests ... and signing off on a diagnosis? AI can do it. But who’s responsible if something goes wrong?, The company that built it? Can we say it’s the doctor who validates the result ... but if I didn’t think it through, didn’t interpret it, didn’t bother making my own neurons work ... do I have to sign it just because the AI produced it?

That’s the contradiction we land in, between what the human and the AI are each supposed to do. And that’s why I’m proposing a different frame, one I’ll try to explain as best I can while I work through what struck me in the paper I read ...

The terrain

To get to that frame, we first need to look closely at the document itself, not to indict it, but to see exactly where the crack I described above opens up.

On August 18, 2026, FDA’s Center for Devices and Radiological Health (through its Digital Health Center of Excellence) published Considerations for the Regulation of Generative AI-Enabled Medical Devices. The first page carries the kind of legal disclaimer only an agency that knows exactly how much weight its words carry would write: this is not guidance, it doesn’t represent a decision, it doesn’t imply legal authority. Formally, it’s an invitation to think out loud.

But invitations from a regulatory agency are rarely innocent. The document opens a public docket (FDA-2026-N-7874) with a deadline: October 19, 2026. Twenty-six discussion questions across five sections, none of which ask whether GenAI belongs in medicine. They ask how, and the how on offer (a two-axis risk framework, a competency-based evaluation modeled on how doctors are trained, a postmarket surveillance system leaning ever more on what happens after approval) isn’t an academic exercise. It’s very likely the conceptual draft of guidance that will, eventually, carry real authority.

Which is why it’s worth reading closely, not because it’s flawless, and not because it needs rejecting. Because in its cracks (in what it takes for granted, in what it leaves as an open question when it should be a design principle) you can read something bigger than the regulation of one kind of software. You can read how this agency, and probably much of the industry behind it, understands what medicine is and where its value actually lives.

Three separate questions start tangling together in the pages that follow, even though the document treats them almost as one: how a tool actually shapes a clinical decision, who can catch a failure and step in on time, and what role we actually want AI to play inside medical reasoning. Let’s start with the first.

How the tool shapes the decision

FDA proposes a two-axis matrix as an organizing heuristic. On one axis, device “activity”: from non-directive information to fully autonomous action. On the other, the “consequence” of an incorrect output: from limited to severe. Risk, they say, climbs diagonally, from the lower-left corner toward the upper-right.

It’s a clean picture. It starts cracking the moment it has to carry the weight of a real clinical interaction.

The document admits as much itself: a conversational device can start out performing an informational function and, over the course of an exchange, migrate into one that directs the user’s action. Here’s how they put it:

A device may begin by performing an informational function and migrate, over the course of a conversation, to an action-directing function. Section IV, “Multi-turn conversations” (p. 9)

Given the paper’s own purpose (inviting discussion, not settling it) it’s consistent that they leave this as discussion question five of twenty-six. The problem isn’t that it stays open. It’s what the matrix itself can’t capture by design: a device classified as “HCP-supervised” says nothing about whether that clinician actually scrutinized the recommendation, or simply accepted it because it arrived wrapped in the fluency and apparent authority of a generative system. The document names this risk elsewhere, it calls it automation bias, the risk of accepting an output without adequate scrutiny because of a system’s fluency or perceived authority, without ever tying it back to its own “activity” axis.

There’s something here that runs against the nature of clinical medicine itself, a discipline that rarely stops to think about software regulation. Medicine doesn’t work through discrete acts that slot neatly into a box. (I know what you’re about to say, “there are checklists for the OR”, but those belong to procedural protocol, not diagnosis.) Medicine works through comprehension of context: the same symptom, the same lab value, the same imaging finding changes meaning (and urgency, and the action that follows) depending on everything else around it. Not even the most elementary binary in clinical practice, sick or not sick, is actually binary. It’s a conclusion that emerges from how the whole picture gets conceptualized, not from reading one isolated data point against a threshold.

A framework that classifies risk by the act, is this informational or directive? supervised or autonomous?, instead of by whether the clinician actually examined what the device handed them, runs a structural risk: regulating the shape of the interaction without verifying what’s actually at stake inside it.

There’s another place where this same problem, classifying form without verifying substance, stops being abstract and becomes technical, measurable: in the body of the text, not a footnote, in the section on measurement and signal-processing functions. This is close to home. An auto-generated summary of findings on a slide, the narrative synthesis of a risk score, a natural-language interpretation of a morphological pattern, functions where a generative system turns a technical output into text the clinician reads as if it were an already-reasoned conclusion.

The document recognizes something important about these functions: even though they produce, on the surface, “non-directive” information, they can be high-risk precisely because:

“...the user typically cannot independently evaluate the basis for the output.” Section IV, measurement and signal-processing functions (p. 7–8)

The user, on their own, can’t verify what that result rests on. A malignancy score, a generated summary over a digitized slide, these present as “information,” not “action.” And yet the person receiving them (even an expert pathologist) usually has no way to audit, in real time, what led the model to that synthesis. The document even ties this observation, in a footnote, to an existing regulatory criterion: the independent-review test within the clinical decision support exclusion.

If the user can’t independently evaluate the basis of the output, on what grounds has that output been treated as mere “information” rather than a form of direction? That question opens onto another, more uncomfortable one: if not even the manufacturer always controls that basis, who does?

Who can actually control its behavior

Two fronts make that question especially hard to answer.

The first: Foundation Model Device Master Files, or MAFs. The document recognizes that much of a GenAI device’s behavior (its refusal tendencies, content policies, safety-critical behaviors) may be controlled not by the medical device manufacturer, but by the developer of the underlying foundation model. FDA’s proposed fix is a voluntary transparency mechanism.

The document is careful to note that an MAF wouldn’t constitute authorization of the underlying model, and that the manufacturer would remain responsible for independently demonstrating the safety and effectiveness of its own device. Formal accountability doesn’t shift. But the useful question isn’t who signs at the end, it’s what information the reviewer, and the manufacturer itself, actually has when building a case on a model whose behavior it doesn’t fully control. If the foundation model developer chooses not to disclose, the manufacturer doesn’t stop being accountable; it’s simply accountable with less visibility into what it’s certifying. And that’s exactly the vulnerability the agency itself admits, bluntly, in discussion question 25:

“model developers may have limited incentive to disclose safety-relevant information.” Section VII.C, Discussion Question 25 (p. 23)

The second front is agentic AI: systems that plan and execute multi-step tasks, use external tools, act in sequence without a human stepping in at every turn. The document gives exactly one example of when an agentic function would fall under its oversight (when the system’s sequence of actions ends up controlling another medical device) without saying where else that criterion might apply. And on functions like care coordination, clinical documentation, or patient outreach, the document only says they “may not be” the focus of its regulatory oversight, a conditional statement, not a categorical exclusion, leaving the door open without committing to walk through it.

The question isn’t whether these functions should fall under FDA’s regulatory umbrella. It’s a question of role. An agentic system can (and probably should) handle summarizing, drafting, organizing. It can propose a diagnostic or therapeutic plan. It can even execute that plan, as long as there’s explicit, verbal confirmation from the responsible physician before any step that matters. What it can’t have is its own initiative inside the diagnostic process: ordering a test on its own, requesting an analysis, acting without that human confirmation at the point where the decision gets made.

Because an agentic AI acting alone, without that checkpoint, isn’t expanding the shelf. It’s closing it. It collapses the diagnostic process into a single path (whichever one the system decided to follow) exactly the opposite of what this technology is supposed to deliver. What we ask of that library, and who decides how it gets used, is exactly the question still left unresolved.

The diagnostic library

This is worth pausing on, because everything above (influence that goes unverified, control diluted between manufacturer and supplier) converges into something bigger than a technical flaw in the risk matrix.

The document proposes evaluating GenAI devices through a “competency-based” approach, explicitly modeled on how physicians are trained and credentialed: knowledge exams, supervised practice with growing independence, ongoing reassessment. The analogy is seductive. It’s also, on one crucial point, misleading.

Citing Freyer et al., the document itself acknowledges:

“human clinicians face professional, legal, and reputational consequences when expected standards are not met.” Section V.A, citing Freyer et al. (p. 10)

Physicians, when they fail the expected standard, face real professional, legal, and reputational consequences. It’s a fact that surfaces, gets cited, and disappears. The framework built afterward never resolves its device-side equivalent: who loses the license when the model gets it wrong? The manufacturer? The foundation model developer who never even touched the design of the final product?

But stopping there (at the accountability gap) sells the problem short. The deeper issue is this: the whole framework, by borrowing the shape of clinical training, risks imagining generative AI as if it were meant to become a substitute for the medical act itself. As if the end goal were for the device to, eventually, do what the doctor does, just faster.

That’s a fundamental misreading of where medicine’s value actually lives.

Much of what goes wrong in medicine isn’t a matter of speed. It’s a matter of blindness: even the best-trained physician can only diagnose what their training taught them to look for, and only treat what their experience lets them recognize. Every clinician works from a diagnostic library (the set of hypotheses, patterns, and possibilities they have on hand to consult when a case walks in) and that library, by definition, has limits. A large share of diagnostic errors aren’t cases of choosing wrong between two options that were actually weighed. They’re cases of never considering the right option, because it wasn’t on the shelf.

That’s where the real value of these tools should sit. And credit where it’s due: the document doesn’t entirely miss this. Appendix A devotes a full element (E.2, information gathering and clinical analysis) to asking whether a device generates complete differentials and avoids premature diagnostic closure, and discussion question 14 explicitly asks when human-AI team performance, rather than the device alone, should be the basis for evaluation.

The problem isn’t that FDA doesn’t see the breadth of the catalog. It’s that it measures the breadth of the output (how many hypotheses the system generates) without establishing whether those hypotheses actually get folded into the physician’s reasoning, or whether they just pile onto a screen the clinician learns to scroll past. Measuring the shelf isn’t the same as measuring whether anyone consults it.

This isn’t about making doctors more productive, that’s a speed metric, cases per hour. It’s about making the diagnostic process more efficient and less error-prone, a completeness metric: how much of the relevant library actually made it onto the table before the decision, and whether that library got used at all. These sound like the same goal. They aren’t.

Under this logic, a GenAI system shouldn’t be designed (or regulated) as if it aspired to replace the physician at the moment of decision. It should be designed as what widens the shelf: what surfaces relevant hypotheses that would otherwise have gone unnoticed, what expands the catalog of the possible before the decision lands with whoever is actually making it, in front of an actual patient.

AI suggests. The physician decides. Not because this is some sentimental hierarchy of “human over machine,” but because there are two distinct accountabilities at play here, and they shouldn’t get confused: the manufacturer answers for the safety and performance of the product it built, that doesn’t go away, and FDA’s own framework exists to enforce it; the physician answers for the clinical decision made in front of that patient, with the information available at that moment. A system can be flawless on the first count and still not exempt the physician from the second. That’s why the final call (not accountability for the product, but for the clinical act) has to stay exactly where it’s always been.

A regulatory framework that doesn’t clearly distinguish between “a device that decides” and “a device that widens what the decision-maker can consider” can’t, as a result, protect what actually matters. It can police substitutes. It cannot demand bigger libraries.

Widening the shelf isn’t free, though. A bigger library can also turn into noise: low-probability hypotheses that pull attention from the real picture, alerts that pile up until the physician stops reading them closely. Prioritizing and filtering are, in fact, a necessary part of any usable library, nobody wants a hundred unsorted hypotheses. The risk isn’t in filtering. It’s in filtering without showing your work: narrowing the catalog to whatever the system is most confident about, without communicating that uncertainty, without tracing each suggestion back to the evidence behind it. That hands the physician a single path dressed up as several options. Calibration isn’t about filtering less, it’s about filtering while showing why.

The questions still open

There’s no clean conclusion here, and there probably shouldn’t be one yet. FDA’s document is, literally, asking for answers, the docket stays open until October 19, and intellectual honesty means admitting that many of these questions don’t have an obvious answer, let alone a single correct one.

How do you regulate a medical act that, by definition, can’t be judged outside its context? What standard of transparency is reasonable to demand from a foundation model developer who currently has no structural incentive to give it to you? Where exactly is the line between “a tool that widens the library” and “a system that starts, without anyone quite noticing, deciding on its own”? And maybe the most uncomfortable question of all: if the digital health industry (myself included) has spent years talking about “clinical productivity” as the metric that justifies investing in these tools, have we been aiming at the wrong target this whole time?

I don’t have the answers. What I do have is a more developed, more rigorously sourced set of observations, which I’m preparing as formal comments to FDA’s docket — because a document like this one, still being written in real time, is exactly the kind of space where it’s worth putting on record the voice of the people doing this work every day, not just commenting on it in a post.

Who Needs Whom? AI and Pathologists: A Love Story

Dr. Luis Cano — Fri, 18 Sep 2026 12:33:15 GMT

The dinner and the question

A new face joined our table at a dinner with friends. He worked outside medicine. We were doing what doctors do, half-joking about whether AI was going to end us all, when he cut in.

AI is already here, he said. You need to accept it, whether you like it or not. It’s staying. Adopt it or become obsolete.

Nobody spoke for a few seconds. Then someone said, half-joking: well, if it’s staying, I hope it pays rent.

The line was silly. It worked anyway, the table laughed, the pause broke, the conversation moved on. But something about how it landed stayed with me on the drive home.

Every conference, every post, every headline runs the same direction: pathology (radiology, medicine in general) needs AI to move forward. I hear the reverse question far less often. Does AI need pathologist to move forward?

“Pathologists” isn’t one negotiating bloc. A hospital can take an equity stake in a company while the pathologists whose archive made that company possible see none of it and have no voice in the deal. A pathologist can annotate thousands of images for a research project without gaining any say in what gets built with that work afterward. Needing the clinical knowledge a profession has accumulated is not the same as needing the collaboration (let alone the negotiated consent) of everyone who spent decades generating it: pathologists, but also lab technicians, and ultimately the patients whose tissue made the archive possible in the first place.

Contributing knowledge to a tool, validating it clinically, and deciding what gets built with it are three different things. They rarely land with the same person.

What clinical practice actually provides

What’s easy to trace is what the industry needs, concretely, to build any of these products. Training a diagnostic model requires whole-slide images tied to clinical diagnoses already confirmed by someone with expert judgment, an archive no company in the sector seems to have assembled without drawing on decades of hospital-by-hospital clinical work.

Exactly how that knowledge gets extracted varies. Some models still depend on extensive manual annotation, delineated case by case. Others, like TITAN, pair self-supervised learning with alignment against the text of existing pathology reports, cutting that dependence without eliminating it. Either way, clinical knowledge remains the input that makes everything else possible.

Two more resources matter, and no architecture fully replaces them. One is validation under real clinical conditions. Paige Prostate, evaluated in the ARTICULATE PRO trial within the NHS, wasn’t validated on lab-bench performance alone; pathologists at Oxford, Bristol, and Coventry reviewed its outputs case by case before the system could scale. The other is credibility. When a company announces a partnership with a leading institution, it’s transferring medical prestige onto a technology product, a signal that lowers the risk an investor perceives.

Not every company depends on this the same way. Owkin’s federated learning model, where images never physically leave the hospital that generated them, keeps legal custody with the institution even as it hands over the computational value that material helps produce, a different architecture from companies built on exclusive access to a single institution’s archive.

A concrete case: who negotiates, who participates

In 2018, Memorial Sloan Kettering Cancer Center gave Paige.AI exclusive access to its archive of 25 million pathology slides, built over decades of work by its anatomic pathology department, work that exists because thousands of patients agreed to let their tissue be studied and kept, not only because a profession generated it. In exchange, the hospital received roughly a 9% equity stake in the company.

That September, ProPublica and The New York Times reported that three people with authority inside the hospital, a member of the board’s executive committee, the chair of the pathology department, and the head of its computational pathology lab, were also Paige.AI co-founders with personal equity stakes. Pathologists in the department, the ones who had generated the archive, not the executives who negotiated the deal , protested publicly. The complaint wasn’t that the hospital received something in return. It was that the people who set the terms of the exchange were, in part, the same people who personally benefited from those terms. Under public pressure, the pathology chair announced he would give up his stake, and the hospital imposed a moratorium on future board-member investments in companies spun out of its own assets.

This is one case, not a sector norm, there’s no evidence this repeats across every deal. But it’s the clearest, best-documented answer to the question that opened this piece: who negotiates on behalf of knowledge that many people, over many years, helped produce, and whether those people had any say in the terms.

What we should be asking when we’re invited to collaborate

An institution receiving equity doesn’t, by itself, prove the exchange was unfair, equity is precisely a way of sharing in future value, not a denial of it. What’s worth asking, case by case, is how the clinical contribution was valued before that percentage was set, who represented the professionals’ interests in that negotiation, and whether they had any influence over what came next, or only over validating it.

That’s maybe the most useful distinction here: annotating a dataset is not the same as deciding what algorithm deserves to be built in the first place. At Memorial Sloan Kettering, the department’s pathologists were involved in generating the archive and, once the deal became public, in protesting it, not in negotiating its terms or shaping what would be built with it. Whether that describes most companies in the sector, we don’t know. It’s probably the most important question this piece leaves open.

PathAI’s trajectory shows what a buyer values. Roche’s May 7, 2026 announcement of its acquisition ($750 million upfront, up to $300 million in milestones) gives roughly equal weight to PathAI’s image management system and lab efficiency as to companion diagnostics for pharma. Capital in this sector rewards an algorithm that can serve two markets at once, clinical and pharmaceutical. Whether that means a pathologist’s everyday needs (better LIS integration, less workflow friction) end up losing out to products with a clearer path to a pharma deal is a fair question. This case alone doesn’t answer it.

Building, maintaining, and getting a tool like this into real clinical use also takes work, capital, and real financial risk, risk almost no hospital is positioned to carry on its own. None of that contradicts what the Memorial Sloan Kettering case shows. It confirms the dependency runs both ways, and that the real question isn’t who needs whom more (that math probably has no clean answer) but whether the terms of the exchange get decided with the people who generated the underlying knowledge at the table, or only with the people who have more experience negotiating deals like this one.

If someone said it to me again, at that same table, that AI is here and I need to accept it or be left behind, I’d go back to the rent joke. This time, I’d mean it.

What do we contribute to make that technology exist? What do we get in return? And what say do we have in what gets built with our work?

We can recognize a tool’s value and keep asking those questions anyway. I’d like them on the table the next time someone hands us the future as something already decided.

Credential Theater: Should a pathologist learn to code?

Dr. Luis Cano — Fri, 11 Sep 2026 09:58:47 GMT

The Perfect Precedent

Someone once asked me: “Dr. Cano, could you teach me how to become a digital pathologist? What do I need to do? Do I need to learn to code?”

With as much humility as I could manage, I said: I’m not a digital pathologist. I’m just a pathologist who got interested in the digital side. I can code and build algorithms because I like computing, and because I have trouble sleeping at night.

I could see, on my listener’s face, a look of broken expectations; he felt lost, again. That same day, in my late-night thinking, I asked myself the question he’d asked me: should a pathologist learn to code? So I decided to look into it. What I found left me with more questions than answers, and a bad taste in my mouth, because what I thought was a training gap turned out to be something harder to name.

For 25 years there has existed a formal subspecialty called Molecular Genetic Pathology (MGP), jointly certified by the American Board of Pathology and the American Board of Medical Genetics and Genomics since 2001. It emerged from a deliberate process, with its own board exam, a standardized curriculum, and the full weight of regulated medical certification behind it. On paper, it’s exactly the precedent the question needs: proof that when science outgrows what a generalist can reasonably handle, medicine doesn’t wait for each individual to reinvent themselves. It builds a new professional box, with rules, with an exam, with a medical board that vouches for its rigor.

The question doesn’t come out of nowhere. It comes from a feeling I share with almost any pathologist who’s been practicing for more than five years: the volume of information a complex case carries today (molecular profile, transcriptomic signature, digitized immunohistochemistry panel, an AI model suggesting a recurrence probability) easily exceeds what traditional training prepared any of us to process in the time a clinical shift allows. That feeling is real, and it doesn’t need to be questioned. What does need questioning is the solution almost everyone proposes in the same breath: if the volume of information grew, the fix must be for the individual to learn more: programming, bioinformatics, applied statistics. It’s an intuitive answer, and that’s exactly why it’s worth being suspicious of it before adopting it.

And this isn’t the only time medicine has resolved a similar tension. Between 1990 and 2005, radiology absorbed the rise of PACS systems and 3D reconstruction without creating a single new subspecialty: the American Board of Radiology simply redefined the required competencies of the primary residency, requiring mastery of digital workstations without ever asking the radiologist to program the software running behind them. Cardiology, between 1995 and 2015, chose a middle path: instead of an ABMS subspecialty, it built a tiered credentialing system (run by independent boards like the National Board of Echocardiography) that let general cardiologists certify in advanced cardiovascular imaging without fragmenting the primary medical title. Two similar technological tensions, two different institutional solutions, both apparently successful. I should make an honest caveat here: both cases come from a general description within the research this piece draws on, without the same chain of verified primary sources that backs the MGP figures. I’m using them as illustrative context that different institutional paths exist beyond creating a brand-new subspecialty, not as evidence carrying the same weight. Even with that caveat, the question they raise still stands: if equally plausible, different methods already exist, why not replicate one of them for digital and computational pathology?

What the Scale Actually Says

It’s worth noting, though, that neither of those two precedents turned the specialist into a programmer. The radiologist who masters PACS doesn’t write the software running behind the system; the cardiologist certified in cardiovascular imaging doesn’t program the reconstruction algorithm their CT scanner uses. In both cases, the institutional solution expanded the specialist’s interpretive competence (what they know how to read, what to order, what to question) without merging their role with that of whoever builds the tool. It’s a distinction that gets lost easily in conversations about digital pathology, where “learning to code” and “learning to interpret what an algorithm produces” tend to get mentioned as if they were the same requirement.

The answer starts by looking at the real scale of the precedent that seems to confirm everything, and that’s where my first certainty began to crack. From the first certification exam in 2001 through the close of the 2024 cycle, the American Board of Pathology has awarded a cumulative total of 882 MGP certifications. I compared that against the rest of the system: more than 58,000 certifications accumulated by the same organization since 1936, 5,088 in Cytopathology, 3,739 in Hematopathology. I want to make one clarification before drawing conclusions: these are cumulative certifications, not necessarily active practitioners today, and comparing subspecialties of different ages and eligible market sizes without adjusting for those variables can impress more than it explains. Even so, the gap (hundreds versus tens of thousands) supports a more limited point: MGP never became the norm of the system.

A national study conducted by the Association for Molecular Pathology together with the ABPath on its own diplomates adds more nuance than isolated figures can. Only 28% of certified diplomates devote all or most of their clinical time to signing out molecular tests; 32% combine it with anatomic pathology, 23% with hematopathology, and 13% don’t sign a single molecular report in their current practice. Read carefully, that distribution doesn’t by itself prove the subspecialty has low demand: the 55% who combine it with another practice may well be the expected working model for a niche specialist, not evidence of structural weakness. What I found harder to square with a healthy labor model is another finding from the same survey: MGP’s own diplomates identified training gaps in their fellowships around next-generation sequencing applied to somatic oncology, the very technology that now defines much of the field’s daily work. Not even those who completed the full certification process feel their training reached the level their job ended up demanding of them.

MGP isn’t, then, proof that medicine easily creates hybrid specialists. It’s proof of something more nuanced: even with a quarter century of serious institutionalization behind it (board exam, standardized curriculum) MGP never became the dominant path. Unlike Nuclear Medicine as a standalone residency (a case I left out of this piece for lack of a primary source solid enough to back it fully, but where the market reportedly ended up favoring dual-certified specialists over the exclusive title) MGP didn’t shrink: it grew steadily, from 30 programs in 2009 to about 58 fifteen years later, with occupancy rates that held between 83% and 95% throughout that period. Calling it “tiny” would be inaccurate; it’s more precise to say it’s small and stable. The mistake isn’t in MGP. It’s in treating it as the mandatory model to replicate for whatever other technological tension arises in the lab, when its own history shows that deep, minority specialization can coexist, without conflict, alongside a general curriculum that absorbs basic interpretation for everyone else.

Two Places Where It Should Show

If the hypothesis I started with were true, that the market is pushing the generalist pathologist toward bioinformatics, and that the logical response is to replicate the MGP model, or radiology’s, or cardiology’s, for the digital and computational layer, you should be able to see it in two very concrete places: mandatory training curricula, and real job postings. Let me show you both, one at a time, because neither fully confirms what my original question assumed.

The curricular frameworks of the UK and Canada are the most explicit that exist outside the United States, and both tell a similar story. The Royal College of Pathologists’ Histopathology Curriculum 2021 made mandatory, within its Capabilities in Practice, both the clinical interpretation of genomic results integrated into the final report and the use of Whole Slide Imaging viewers for routine diagnosis. In Canada, the shift was even more symbolic: in July 2023, the Royal College of Physicians and Surgeons officially renamed the Anatomic Pathology specialty to Diagnostic and Molecular Pathology, folding genomics directly into the primary title without creating a separate secondary board. In both countries, it’s institutional recognition that interpreting molecular results and navigating a digital viewer are no longer extras: they’re part of the base craft of any pathologist trained today.

In Mexico and Brazil, the picture is more uneven but points in the same direction. UNAM’s Anatomic Pathology specialty program and the criteria of the Consejo Mexicano de Médicos Anatomopatólogos treat molecular and digital pathology at a basic interpretive level, with no national curricular framework making it uniformly mandatory. In Brazil, USP’s Residência Médica em Patologia leaves exposure to bioinformatics and digital pathology to elective rotations, depending on each training center’s infrastructure. Neither country requires programming as a competency. Neither resolves the problem with the same force as the UK or Canada either: it simply leaves it unresolved, scattered across individual programs.

It’s worth pausing for a moment on how that “mandatory” gets measured, because the evaluation method reveals as much as the content. The UK doesn’t count classroom hours: it evaluates through Supervised Learning Events during real practice and an annual review of progress, the resident demonstrates competency by working, not by memorizing a syllabus. Canada built a similar system with its Competence by Design model, where the resident moves through stages, Transition to Discipline, Foundations, Core, Transition to Practice, and must reach specific milestones, like Core-stage milestone C8, which requires applying knowledge of human genome structure and the principles of analytic validation for genetic tests before advancing. The documents I reviewed in both systems don’t include programming as a required competency at any evaluable milestone, a consistent absence across two different health systems, with two different evaluation philosophies, arriving at the same place without needing to coordinate with each other.

And when I reviewed real job postings aimed at generalist pathologists (not MGP subspecialists, but those signing out diagnoses in community practice) the technological requirement that kept showing up was operating a Whole Slide Imaging viewer to diagnose on a screen instead of an optical microscope. In the sample I reviewed, I didn’t find a single generalist pathologist posting that asked for Python, data analysis, or bioinformatics pipeline development as a requirement. I want to be precise about the weight of this evidence: the sample is small, and much of the salary data I found came from secondary sources (specialty discussion forums) rather than formal salary surveys. With that caveat, what I can say with confidence is more modest than “the market doesn’t ask for it”: in the evidence I managed to gather, I didn’t find that requirement. The market, as far as I could verify, isn’t asking the pathologist to become a data engineer. It’s asking them to know how to read, on a screen, what they used to read under a microscope.

One of the postings I reviewed illustrates the pattern well: a remote listing for “General Pathologist – Digital Pathology” from a diagnostic services company, aimed at pathologists with standard AP or AP/CP certification, with no MGP required. The job description asks for primary histopathologic diagnosis via Whole Slide Imaging, reading the case on screen, using the viewer’s zoom and annotation tools, as a central operational part of the daily work. Nowhere in the description does a word appear about programming, statistical analysis, or algorithm development. It’s the kind of posting that, read quickly, seems to confirm that “digital pathology is already here” and that the pathologist must adapt at breakneck speed. Read carefully, it confirms something more modest and more precise: the pathologist has to adapt to a work tool, not to a second profession.

Consider the most common hypothetical case in any reference lab today: a lung biopsy with suspected adenocarcinoma. The pathologist examines it on a digital viewer, not under a microscope. The report they must sign integrates, alongside the histologic diagnosis, a next-generation sequencing panel with half a dozen actionable genes and a PD-L1 expression score; in a growing number of labs, the output of an algorithm estimating the probability of a specific mutation from the image itself gets added too. The pathologist didn’t build the NGS panel, didn’t train the algorithm, didn’t write the pipeline that aligned the sequences. Their job is different: deciding whether that data integration, taken together, supports a defensible clinical conclusion, and signing that conclusion with their name and their license. That’s exactly the kind of task the UK and Canadian curricula already recognize as a required competency. And it’s also exactly the kind of task no amount of Python classes solves better than a well-designed curriculum for integrated interpretation.

The Question That Loses Its Footing

So the question I started with “should a pathologist learn to code?”, doesn’t have, at this point, a universally affirmative answer. In the curricula and job postings I reviewed, programming doesn’t appear as a core competency for the generalist pathologist: the curriculum doesn’t require it, in any of the countries I reviewed in detail; the market doesn’t require it, at least not in the sample of postings I analyzed. That doesn’t mean programming is worthless (it can still be useful for specific academic, industry, or research roles) but it does mean the premise I started with, that of a generalized requirement, doesn’t hold up. And the precedent that seemed to confirm it, MGP, turns out to be a niche subspecialty that most of its own diplomates don’t even practice full-time. If you’ve made it this far expecting an answer about individual skills, the ground under that universal question just gave way. What’s left is a different, more uncomfortable question: if no one is formally requiring this transformation, why does the sense of pressure persist so strongly among young pathologists, program directors, and the industry that funds them?

Credential Theater

One possible hypothesis is that the pressure doesn’t come from a demand for technical skills that’s simply been misframed, it comes from somewhere more uncomfortable: a gap between the vocabulary of institutional legitimacy and the actual institution backing it. Today, no medical board, not the American Board of Pathology, not the American Board of Medical Specialties as a whole, independently certifies a subspecialty in “digital pathology” or “computational pathology.” The only regulated medical certification that partially covers that territory remains Clinical Informatics, and only 252 diplomates coming specifically from pathology have earned it since it became available, in 2013.

What exists instead is a map of four recognition categories that rarely get distinguished from one another in everyday language:

Four categories, four different levels of real weight, and a single term, “fellowship,” that names all of them equally in public conversation.

The sociology of work literature describes the risk at play in that distinction with two words that rarely make it into conversations among pathologists: deskilling and upskilling. The deskilling risk is that automating microscopic scrutiny reduces the pathologist’s perceptual sharpness, turning them into a passive validator of whatever the algorithm already decided. The upskilling opportunity is the opposite: that digital viewers and integrated genomic reports free the pathologist from repetitive tasks (manually counting mitoses, hand-scoring the Ki-67 index) and let them focus on the synthetic integration of complex cases, which is, in the end, the part of the craft whose clinical responsibility still belongs to the pathologist. Which of the two scenarios wins out doesn’t depend on how much Python the pathologist learns. It depends on whether the institution that trains and employs them seriously builds the intermediate layer that lets them use the tool without becoming its hostage.

And yet, each of the non-accredited programs advertises itself using the same language of legitimacy as an accredited MGP fellowship. “Fellowship in computational pathology,” says the cancer center’s page. “Accredited program,” says the institutional brochure, even when the accreditation in question comes from the university itself and not from ACGME. It’s a distinction that neither the applying resident nor the VP of medical affairs funding the program tends to notice, because the word “fellowship” already carries, by habit, a promise of rigor that in MGP is genuinely guaranteed by a board exam, and in digital pathology, simply isn’t.

I call this Credential Theater (and I’m presenting it as a question, not a verdict) : are we looking at real institutionalization of digital and computational pathology, or are we using the vocabulary of a subspecialty before building the mechanisms that would give it that weight? The evidence I gathered doesn’t include surveys of fellows, interviews with program directors, or documented cases that would turn suspicion into a closed accusation. What I do have is the objective gap described above, and the framework sociologist Andrew Abbott used to describe how regulated professions negotiate control over complex tasks: a resolved jurisdictional settlement clearly divides up who commands, who executes, who advises, among the professions competing for the same task. Digital and computational pathology doesn’t show, as far as I could verify, that resolved division. It shows, for now, the vocabulary of one.

In the diagnostic lab, three distinct professional groups coexist, hardly ever named out loud, around the same territory of tasks: the medical pathologist, who holds legal jurisdiction over tissue interpretation and therapeutic decisions; the bioinformatician or data scientist, who masters the algorithms and the sequencing but holds no medical license and no direct responsibility for the patient; and the lab technician, who carries out the analytic processing and the scanning of slides. When the process works (as it seems to have worked in radiology, as in cardiology) what emerges isn’t a merger of these roles into a single professional of impossible breadth, but what the organizational management literature calls a T-shaped profile: vertical depth in diagnosis and the legal responsibility of sign-out, horizontal breadth enough to interact with the bioinformatician and interpret their alerts without needing to build the tool. The problem with digital pathology, then, doesn’t seem to be that its practitioners lack that T-shaped profile. It seems to be that it lacks the institution to define it, train it, and certify it with the same seriousness MGP used to define, train, and certify its own territory 25 years ago, though, again, that last part is my hypothesis, not a demonstrated conclusion.

What We Still Don’t Know

Industry actively participates in this scenario, with real budgets behind it: scientific societies like ESMO offer digital and computational pathology fellowships funded directly by pharmaceutical companies, ESMO itself runs one such program funded by AstraZeneca. But beyond that, honesty requires admitting I have more questions than answers.

What do candidates actually think they’ll get out of enrolling in one of these fellowships? Do they understand the difference between an ACGME-accredited program and an institutional finishing diploma, or do they reasonably assume the word “fellowship” already guarantees something? How do hiring committees, inside and outside academia, actually value these credentials? A fellowship not recognized by any medical board can still carry considerable job-market value (through the center’s prestige, the skills acquired, the professional network it leaves behind) and I have no evidence the market is discounting it. And what does a funder (a pharmaceutical company, a scientific society) mean by “success” when it approves the budget for one of these programs? Publications? Graduates placed in industry? Future formal recognition? I found no public criteria that would let me know how funders define success for these programs.

None of these three questions has, as yet, an answer I can document with the same solidity as the MGP figures. That’s exactly what turns Credential Theater into a hypothesis rather than a conclusion: it’s testable (through surveys of fellows, interviews with program directors, analysis of promotional materials) but no one has tested it yet. Neither have I.

The Questions a Candidate Should Ask

Picture a resident comparing two digital pathology fellowships, side by side. Can they tell, just from reading the brochures, which one leads to a certification recognized by a medical board and which one is an institutional finishing diploma? Does the acceptance letter explain that difference? Do they know how their next employer will value that year of training three or five years from now? I don’t have, yet, data to answer those questions in general terms. What I can offer is the list of questions that, based on what I found in this research, I’d suggest asking any institution before signing:

Who accredits this program, a university, a medical center, or a medical board recognized by ACGME?
Does it lead to eligibility for a subspecialty exam, or is it a finishing diploma without that path?
What clinical functions will I be able to perform after graduating that I couldn’t before starting?
Where do the program’s past fellows work now, and in what kind of role?
How do employers describe this training when they see it on a résumé?
What competencies are formally evaluated during the program, and who is accountable for that evaluation?

I don’t know how often candidates, institutions, and funders ask these questions before committing. The lack of comparable information is exactly what makes answering them unnecessarily hard.

I don’t know, precisely, how many residents are in this situation right now. Unlike MGP, which has a public, auditable record of its 58 active programs, digital and computational pathology fellowships have no equivalent census, neither the Pathology Informatics Association nor any other organization keeps a full count of how many exist or how many residents pass through them each year. That absence of a registry is itself part of the problem: it’s not just accreditation that’s missing, it’s the very ability to measure how many people its absence affects.

The Right Question

The question I raised a while back “should a pathologist learn to code?”, turned out to be the wrong one. Not because it doesn’t matter, but because the system already resolved it without anyone announcing it: the curriculum requires interpretation, not code; the market requires a viewer, not a programming language. The right question is a different one, and it should make both universities and the industry funding them uncomfortable: when an institution calls something an “accredited fellowship” for a program that doesn’t lead to a medical certification, what exactly is accredited, and what should the candidate signing up understand?

Let me close by returning to the question that opened this piece, the one from the reader building curricula using these essays as reference. If that person is designing a digital pathology training program for their university right now, the real usefulness of this essay isn’t telling them whether or not to include a programming module. It’s giving them a different question to ask every institution they negotiate a partnership, funding, or accreditation with: as designed, will this credential be recognized by something more than the good faith of whoever grants it?

The diagnosis no one can sign: Digital Pathology's global blind spot

Fri, 04 Sep 2026 09:56:38 GMT

A patient once arrived, during my residency, carrying his own histology slides for review. His home hospital had given him a diagnosis that raised more doubts than certainties. The decision to seek an expert’s second opinion was immediate. That first diagnosis was already a month old, a month he had spent raising money for the trip. He waited two more weeks for the review. The trip home would take a couple of days, and a new appointment with his treating physician would take, at best, another week or two. Almost two months.

Two months without a diagnosis. Two months without treatment. Two months of uncertainty in an oncology case.

I wrote about that memory a few months ago, in a piece on telepathology for Roche Diagnostics. My conclusion at the time seemed clear enough: whole-slide scanning and telepathology networks remove the need for that trip. Distance stops being the obstacle.

Last week I came back to that memory for a different reason. I was reading a study just published in Nature Communications on ALLocate, a low-cost AI-powered microscope for acute leukemia diagnosis. The system attaches a 3D-printed scanner to a standard microscope, costs $150, and reaches 88% accuracy on physical slides, against commercial scanners that run past $300,000. The authors frame it, fairly, as a step toward “bridging the gap between AI innovation and practical deployment in resource-limited settings.”

It’s a beautiful chart. $300,000 on the left, $150 on the right. The kind of comparison that gets applause at a conference and headlines in trade press.

And yet, once I’d read the full paper, I wasn’t thinking about hardware cost. I was thinking about that patient. Because there’s a line, near the end of the discussion section, that quietly reframes everything the chart is promising.

A decade solving the same problem

ALLocate isn’t an isolated case. It’s the latest link in a chain of at least a decade of engineering work aimed, explicitly, at solving the cost of diagnostic microscopy.

In 2019, Manu Prakash’s team at Stanford introduced Octopi, a low-cost automated microscopy platform for infectious disease. Its successor, Octopi 2.0, reached sensitivity and specificity above 97% detecting malaria in a clinical cohort of 73 patients. In 2023, a Spanish team presented iMAGING, a similarly low-cost, universal robotic system for the same purpose. In 2024, a University of Chicago group published an open-source workstation (built on an OpenFlexure microscope and a Raspberry Pi) capable of capturing histology images for $230, with diagnostic performance (AUROC 0.80–1.0 depending on cancer type) matching high-end commercial scanners.

Four projects, four teams, four countries, six-plus years of work. They don’t do exactly the same thing (different samples, different levels of validation, different degrees of real-world deployment) but they all point the same direction: hardware cost is no longer a valid excuse to deny quality diagnosis to a low-resource setting.

And on that point, with real rigor and cross-validation, they’re right: certain image capture and classification tasks can be done with extraordinarily cheap hardware. What these projects don’t yet show (and it’s a distinction worth holding onto) is that this hardware can replace the full infrastructure, workflow, and clinical liability of a digital pathology service. A $150 device bolted to a microscope is not functionally equivalent to a $300,000 commercial scanner; it solves one specific link in the chain, not the whole chain.

Stop at the first half of that sentence, and the conclusion looks obvious: access to digital pathology in middle-income countries is, fundamentally, an image-capture infrastructure cost problem. And thanks to a remarkable decade of work, that problem is largely solved.

The line that changes everything

Back to ALLocate. In its discussion section, after presenting its accuracy numbers, the authors describe how they envision real-world deployment:

“Results can be relayed to a qualified expert pathologist for verification.”

Yan et al., Nat Commun, 2026

It’s a reasonable line, even a responsible one. The authors themselves insist ALLocate is a triage tool, not a replacement for human diagnosis. It’s exactly what a well-designed medical AI system should say about itself.

But sit with the word “qualified.” In the same paragraph where the paper justifies its own existence, it cites a figure: between 0.1 and 1.3 hematopathologists per million people in sub-Saharan Africa. That’s where the logic starts to crack. The system solves the cost of the scanner. But its output (the data point that decides whether a patient has leukemia) still needs, to carry clinical weight, exactly the resource the paper’s own premise says doesn’t exist.

This isn’t only an African problem. It hits middle-income countries like Peru with a different, arguably more uncomfortable logic, countries that don’t fit the “poor nation needs charity” narrative, but face the same structural gap.

Peru has 154 registered pathologists. 47% (72) are concentrated in Lima, regions like Amazonas, Áncash, Huancavelica, and Madre de Dios have none.
La República / Salud 360 study, 2025

Peru doesn’t need only a $150 device to scan slides. It has a public health system, decent tertiary hospitals, and pathology residency programs that have run for decades. A cheap device would probably help in some regions, this essay isn’t arguing the technology is useless. But on its own it doesn’t fix what’s actually missing: enough trained people, distributed where the patients are, not just where the infrastructure already sits. Cutting hardware cost, by itself, doesn’t move a single pathologist from Lima to Madre de Dios.

Not just Peru, and the obstacle isn’t technical

The pattern repeats, with local variations, across the region. In Argentina, the Sociedad Argentina de Patología reports most of its specialists concentrated in Buenos Aires City and Buenos Aires Province, leaving the rest of the country splitting a minority share. In Mexico, there isn’t even an updated official geographic census of active pathologists; the Consejo Mexicano de Anatomía Patológica can only offer a cumulative estimate of graduates from the national health system.

Faced with that disparity, the obvious answer (the one any diagnostics executive would propose in a strategy meeting) is to send the slide to whoever can read it. A pathologist in Buenos Aires reads for an uncovered province. One in Lima reads for Loreto. If needed, one in Boston or Madrid reads for either. The technology to transmit a whole-slide image anywhere on the planet has existed, stably and validated, for over a decade.

And yet that isn’t happening at scale in the region. Not for lack of bandwidth. Because a diagnosis isn’t just an image, it’s a medical act, and medical acts have jurisdiction.

Colombia regulates telemedicine and electronic-signature validity under Resolution 2654 of 2019, requiring the provider to be registered in the national health registry (REPS); whether that covers an individual professional acting from another jurisdiction, or only the institution, is a distinction the rule doesn’t clearly resolve, and one that deserves specialized legal reading, not mine. Peru folds pathology into its telehealth framework (Law 30421, regulated by Supreme Decree 005-2021-SA), requiring registration with the Peruvian Medical Association and an authorized host institution. Argentina enables nationwide medical teleassistance under Law 27,553, with digital signatures under Law 25,506. The national statute itself doesn’t explicitly require the professional to be licensed in the province where the patient is located; any additional territorial requirement would need to be verified against provincial and professional-college rules.

None of these three rules bans telepathology or makes it illegal. What does seem true, from the evidence I could gather, is that none of them offers a clearly documented path for a pathologist licensed exclusively in one country to take formal, signed responsibility for a primary diagnosis on a patient in another. The difficulty spans licensing, institutional authorization, data protection, and professional liability and its exact scope varies by jurisdiction. A stronger claim than that would require the kind of comparative legal audit this essay doesn’t attempt.

The pattern has a name, and we’ve seen it before

What’s happening here is a variant of something I’ve called, in this column, Institutional Reductionism: an industry’s (or an entire field’s) tendency to measure its own progress by whatever variable is easy to photograph, fund, and communicate, instead of the variable that actually determines whether the patient gets what they need. $300,000 to $150 is a photographable variable. “Who signs the diagnosis, and where are they allowed to do it?” is not. Which is why one becomes the headline of a decade of brilliant engineering papers, and the other barely gets a mention, almost in passing, in a discussion section.

The most-cited case of successful telepathology in a setting of extreme specialist scarcity is probably Butaro District Hospital, in Rwanda, developed with Partners In Health, Dana-Farber, and Brigham and Women’s Hospital. The program, documented by Mpunga and colleagues, reached diagnostic concordance above 95% comparing static-image telepathology against conventional histology.

It’s rightly one of the most-cited global health programs of the last decade. The published evidence I could locate documents, over its first nine months, a triage system built on static images: it selects which cases deserve priority review, not a model of remote primary diagnosis with liability fully transferred to the foreign specialist. I don’t know whether the program later evolved into something else (the study documenting that initial phase can neither confirm nor rule that out) but even in its most consolidated, most-cited form, that’s the model that proved out. And if that’s the ceiling reached by the world’s most successful case, backed by a decade of philanthropic and institutional investment, the question stops being purely technical.

Latin America has its own attempt at answering the same question, and it’s worth a close look because it doesn’t come from an international NGO, it comes from Colombian pathologists solving their own problem. It’s called GLORIA (Globalización de telepatología en Red con aplicaciones de Inteligencia Artificial), and it aims to connect general labs in small and mid-sized cities with oncology referral centers across Colombia.

GLORIA, like Butaro, is designed, for now, as a support network for complex cases, not a mature national network for remote primary diagnosis, and it doesn’t resolve the full legal architecture of who signs what. What’s notable (what makes GLORIA the most honest piece of this story) is that its own designers say so plainly, in a conference presentation:

“The TP network cannot directly address the shortage of pathologists in Colombia.” Guerron-Gomez et al., Cancer Epidemiol Biomarkers Prev, 2024

The team that built the most ambitious digital pathology project in the region, publishing in peer-reviewed journals with real institutional funding, lands on the same conclusion as this essay: network architecture improves the distribution of complex cases toward the specialists who already exist. It doesn’t increase the number of specialists. It doesn’t change who can sign what, or from where. As of this writing, GLORIA is still in pilot implementation: a baseline study published this year surveyed 57 Colombian pathologists to assess what they’d need for the network to work (13). The road is still being built; no one has yet resolved who’s allowed to drive on it from another jurisdiction.

Two caveats before jumping to conclusions

It would be easy, at this point, to turn this into fatalism: technology is useless if regulation and human capital don’t move at the same pace. That’s not my argument, and recent evidence qualifies it in two directions.

First:
Digitization doesn’t have to be all-or-nothing, a high-volume lab in northeastern Brazil, affiliated with the Universidade Federal do Ceará, digitized 60% of its routine caseload using a mid-range scanner and structured local storage, reaching 98.72% concordance between digital and traditional diagnosis (14). They didn’t digitize everything. They didn’t need a national-network budget. They digitized what made sense to digitize, with infrastructure they could actually sustain, and that already generated real clinical value. It’s a genuine middle path between “everything” and “nothing,” and too many conversations about digital pathology in the region ignore it.
Second:
“There’s a human checking it” can’t be an empty formula. The most documented case in recent literature on this doesn’t come from pathology: it’s the Epic Sepsis Model, a widely deployed prediction algorithm in U.S. hospitals, whose external validation found performance far below what the vendor reported (it missed 67% of real sepsis cases) and yet didn’t cause the disaster those numbers suggest, because physicians kept independently evaluating every patient instead of deferring to the algorithm (15). The lesson isn’t that AI is useless without human oversight. It’s that oversight only protects anyone if it remains the real decision point, with someone identifiable and qualified exercising it, exactly the variable that, in the case of low-cost digital pathology in the region, still isn’t resolved.

The question the industry still isn’t asking

Butaro reached, in its documented nine-month phase, a triage model that worked well, backed by significant institutional and philanthropic collaboration from Partners In Health, Dana-Farber, and Brigham and Women’s, not a system of remote primary diagnosis. GLORIA, in Colombia, is still building the infrastructure to attempt something similar, and its own authors already warn the network alone won’t solve the specialist shortage. Meanwhile, a new low-cost microscope shows up every year (Octopi, iMAGING, Choudhury’s workstation, now ALLocate) solving, over and over, the same variable: hardware price.

I didn’t find, in the sources I reviewed, a Latin American proposal built specifically for cross-border recognition of anatomic pathology diagnosis, something along the lines of what already exists in other regulated professions crossing borders within a shared regional bloc. That doesn’t prove no one has thought about it; only that, within the scope of this research, I couldn’t find one.

But the absence of a Latin American model doesn’t mean this is an exclusively Latin American problem. The European Union has built, over two decades, a system for recognizing medical qualifications based on harmonized minimum requirements, and has determined that, in telemedicine, care is deemed to be provided in the Member State where the provider is established. Those are two essential components of a possible solution.

But recognizing that a doctor can relocate and practice in another country isn’t the same as letting their diagnostic authority travel while they stay put. The European framework does cover temporary cross-border provision of professional services (in some cases a prior declaration is enough) but it typically requires documentation, liability insurance, and, depending on the profession, additional checks by the receiving state. A German pathologist can, under that framework, relocate and practice in France. Far less clear is whether they can open a digital slide from Berlin, issue a primary diagnosis for a hospital in Paris, and have it entered into that patient’s clinical record with an unambiguous chain of liability, insurance, and oversight without ever crossing the border themselves.

The same fragmentation shows up, in different forms, in North America. In the United States, every state medical board requires the physician to be licensed in the state where the patient is located at the time of the telemedicine act, the practice of medicine, the federal reference standard says literally, “occurs where the patient is located”, with narrow exceptions for physician-to-physician consultation or pre-referral screening. Canada holds to the same principle at the provincial level: its national telemedicine framework recommends that the professional comply with both their own jurisdiction’s rules and those of wherever the patient is located. The geography changes (countries, provinces, states) the institutional problem stays the same.

Latin America, then, isn’t the only place missing this infrastructure. It’s the place where the consequences are most visible: few specialists, vast distances, and resources no single hospital can afford to duplicate. It works as a case study of a problem digitization made technically solvable, but that no system (not even the most integrated one on the planet) has made institutionally complete.

The name for what’s missing is diagnostic portability: the ability for a qualified specialist’s clinical authority to cross a digital border without the specialist having to cross it themselves, through a framework recognized by both jurisdictions that resolves licensing, liability, data protection, quality, and reimbursement as one system, not a pile of separate obstacles. Butaro doesn’t have it. GLORIA doesn’t have it. Neither do Germany and France, nor the United States across its own states, nor Canada across its provinces. The image already knows how to travel alone. Diagnostic portability is the name for what we haven’t yet built so authority can travel with it.

I’ve been turning over what a model like that would look like for Latin America specifically, what each professional college would have to give up, how shared legal liability would be structured, what role a regional body like the Andean Community or the Pacific Alliance might play, and I already have a first sketch of how it could work in practice.

Next time someone shows you a bar chart with $300,000 on the left and $150 on the right, it’s worth asking a question before applauding: who’s going to sign the result, and from where are they allowed to do it? The image can already travel. The question (in Latin America, but also in far more integrated systems) is what would need to be built for diagnostic responsibility to travel with it.

P.S.: If your organization is seriously thinking through diagnostic access strategy in the region and wants to explore this sketch further, get in touch.

The Dark Genome and the Two Clocks of Innovation

Dr. Luis Cano — Fri, 28 Aug 2026 09:58:35 GMT

The First Slide

I remember the first time I saw one, though I didn’t know its name yet. It was my first year as a pathology resident. Like every young doctor, I still carried that naive certainty that medicine can cure everyone. He was the one who showed me otherwise.

I was on frozen-section duty (what we called spending the day waiting for calls from the OR for a rapid intraoperative read) when tissue came in from a neurosurgery case. I reached to process it the usual way. I was told not to: call Dr. Casavilca, the staff neuropathologist. He asked for a smear, prepared the slide, looked through the scope, and dictated: consistent with high-grade astrocytoma, suggestive of glioblastoma. Glio, I thought, not yet knowing what that syllable meant. He asked the patient’s age. Thirty-five, the OR order said. What a shame, he said. I didn’t fully understand that line until days later.

The same case came back up at that week’s slide review. The doctor asked if I’d looked at it yet. Embarrassed, I said no. You won’t forget this one, he told me, and he was right: necrosis, hemorrhage, a cellular heterogeneity that looked almost deliberate, the picture of a biological architecture completely wrecked. What that slide suggested, in the language of that moment, was a high-grade diffuse glioma consistent with glioblastoma; a molecular diagnosis that an intraoperative read alone can’t establish, and one that, depending on the year, might have been classified under different criteria than today’s.

What I also never forgot was what the doctor added afterward: that this was one of the deadliest tumors that exist, that it tends to recur within months, that most patients don’t make it to a year. I’d later learn that median survival for glioblastoma runs a bit past a year (12 to 18 months, depending on the population and treatment) but also that no median predicts the individual fate of the person sitting in front of you.

Why Glioblastoma Wins

Under any definition, historical or current, glioblastoma is the most aggressive primary tumor of the adult central nervous system (it occurs in children too, but that’s another story). Much of that aggression comes from a set of traits that work against nearly any therapy thrown at it.

The first is genetic. Glioblastoma tends to carry a relatively low burden of somatic mutations, which limits the supply of private neoantigens (the raw material behind most modern immunotherapies) available for the immune system to learn to recognize. That doesn’t mean the tumor has no immune targets at all: it also expresses tumor-associated antigens and, as we’ll see, antigens potentially derived from the non-coding genome. But it does mean that the most common route in immuno-oncology (chasing mutations unique to each patient) pays off less here than in tumors with a high mutational burden.

On top of that sits a heterogeneity that operates on several levels at once. Intratumoral heterogeneity means a single tumor harbors distinct molecular subpopulations, EGFR-amplified subclones living alongside others driven by PDGFRA alterations, within the same patient. Spatial and temporal heterogeneity matter just as much: the genomic profile of a recurrence, after standard maximal resection followed by radiation and temozolomide, often looks nothing like the original lesion. Under the selective pressure of temozolomide, clones with DNA repair defects frequently emerge, feeding a constant cycle of therapeutic escape.

The tumor microenvironment adds a third layer of resistance. The stroma is typically dense with tumor-associated macrophages carrying predominantly immunosuppressive transcriptional programs (the classic M1/M2 split is, today, a useful shorthand more than an exact classification) and with myeloid-derived suppressor cells, while cytotoxic T-cell infiltration stays scarce. The blood-brain barrier isn’t a simple wall either: its integrity varies across the tumor, and the immune problem it poses runs deeper than physical penetration. Then there’s the standard use of corticosteroids like dexamethasone to manage cerebral edema, which drives apoptosis in naive and memory T cells, one more obstacle, unintended, that standard treatment stacks against any vaccination attempt.

The Dark-Genome Hypothesis

Against that backdrop, several research groups (including the one behind EVX-05) are asking a different question. Not how to train the immune system to fight a tumor that’s already large and entrenched, with every defense described above already deployed. But whether there’s a class of antigen the tumor expresses more consistently across patients than private neoantigens do, one that would let you manufacture a shared vaccine in advance, instead of a custom product built after each diagnosis.

That’s where the dark genome comes in. Roughly 8% of human DNA is made up of endogenous retroviruses (ERVs): fossilized viral fragments that integrated into our genome millions of years ago and that, in healthy tissue, stay silenced through DNA methylation and repressive histone marks. In glioblastoma, global epigenetic dysregulation reactivates many of these sequences, producing proteins that, in a fair number of cases, the adult immune system has never learned to tolerate. That said, this shouldn’t be overgeneralized: not every ERV-derived product necessarily escapes central or peripheral tolerance, and showing that a locus is transcribed is not the same as showing it’s translated, processed, and presented on HLA. That full chain, from transcription to confirmed presentation, is precisely what this essay is about.

It’s an attractive hypothesis for a concrete reason: if certain ERV elements reactivate consistently across different glioblastoma patients, unlike private neoantigens, which are unique to each tumor, it opens a door that personalized immuno-oncology can’t offer at the same speed: an off-the-shelf vaccine, manufactured in advance, with no weeks-long wait for an individualized product while the tumor keeps growing.

Evaxion A/S built its AI-Immunology™ platform on this hypothesis. Based on the company’s public description, which doesn’t allow full reconstruction of the architecture or independent validation of every filter, the system combines several specialized modules. PIONEER™ predicts neoantigens from point somatic mutations in the tumor exome, the conventional approach also explored by platforms like BioNTech’s, Moderna’s, or Gritstone’s. ObsERV™ is the module specific to EVX-05: it scans the non-coding transcriptome to identify reactivated endogenous retrovirus loci, evaluates how often they recur across cohorts of primary tumors, and, per the same corporate description, by comparison against healthy-tissue expression panels like GTEx, screens out any element showing baseline activity in vital organs or healthy brain parenchyma. A third module, EvaxMHC™, then computes the presentation affinity of the resulting epitopes against the most common HLA supertypes in the human population.

On paper, it’s a sensible choice of terrain: where canonical neoantigen predictors underperform in low-mutational-burden tumors, ObsERV™ would sidestep that bottleneck by working on genomic territory that doesn’t depend on how many mutations any given patient happens to have.

What EVX-05 Claims, and What’s Been Shown

On April 17, 2026, Evaxion released the computational data that its team, together with Duke University School of Medicine and under Professor Mustafa Khasraw, would present days later in a poster at the AACR annual meeting: conserved, overexpressed ERV sequences in human glioblastoma samples, absent from reference healthy-tissue panels. Worth flagging in passing, something that would deserve clarification in a future review: the original abstract title referred to a personalized vaccine strategy, while the product was later presented as off-the-shelf. This may simply reflect a normal evolution from early academic work to the final commercial candidate, but the company hasn’t explained it publicly.

So far, this reads the way a well-grounded innovation story should: a well-diagnosed structural problem, a plausible biological hypothesis, a computational platform exploring it with reasonable methodological rigor. The question that matters isn’t whether the hypothesis is interesting, it is. It’s how much of it has actually been confirmed biologically, beyond the computational prediction.

What the public evidence actually supports, step by step:

Of the twelve rungs on a standard preclinical evidence ladder (from biological hypothesis to clinical data in humans) EVX-05 has, as of this writing, three demonstrated or partially supported: ERV re-expression in tumor biopsies, its relative absence in healthy tissue, and computational identification of candidate sequences. Rungs 4 and 5 (actual HLA presentation and immunopeptidomics, recognition by human T cells) are computationally predicted, not experimentally confirmed. The remaining seven (in vitro immunogenicity, in vivo efficacy in animal models, preclinical safety, and any clinical data in humans) have no documented public evidence. The Phase 1 trial planned with Duke University has not begun. (Own reconstruction based on the sources listed at the end of this essay, following a standard twelve-rung preclinical evidence ladder.)

No published mass spectrometry confirms that the ERV peptides predicted by ObsERV™ are actually processed and presented on the surface of a living glioblastoma cell. No orthotopic mouse model has shown that the induced immune response can cross the immunosuppressive microenvironment and blood-brain barrier described above. The hypothesis is biologically plausible and worth testing. The validation that would let us call it solid doesn’t exist yet in the public domain.

Same Day: EVX-05 In, EVX-03 Out

On August 17, 2026, four months after that presentation, Evaxion filed a Form 6-K with the SEC carrying two announcements in the same document, on the same day. First: the formal addition of EVX-05 (a program code, not yet a commercial name) to its active portfolio. Second: the immediate discontinuation of EVX-03, a personalized DNA vaccine for solid tumors the company had been projecting toward clinical trials since its 2021 annual reports.

The company described the decision as a matter of “portfolio management.” Beyond that phrase, we don’t know exactly why EVX-03 was discontinued: its 20-F filings from 2021 through 2024 consistently described it as a mature program headed for Phase 1/2a, and that trial never dosed a single patient, but no public disclosure explains whether the reason was an unresolvable manufacturing barrier for personalized DNA, a preclinical immunogenicity problem, or a capital-allocation call. The timing next to the EVX-05 launch, and the company’s insistence that the restructuring doesn’t touch its cash runway through the second half of 2027, support a financial reading of the decision. They don’t prove one.

According to Evaxion, Duke University is expected to take part in the initial clinical evaluation of the program. The public record doesn’t specify how the costs of that trial are split between the two institutions, it would be a mistake to infer, from the mere existence of the academic partnership, that Duke is absorbing a specific financial burden.

The Precedents That Show Where Prediction Fails

Before drawing conclusions about EVX-05, it helps to see where it stands next to other programs chasing the same goal by different routes, because glioblastoma, for lack of effective therapies, has become a crowded proving ground for nearly every vaccine architecture imaginable, and several of these programs have already hit obstacles EVX-05 hasn’t faced yet.

At the personalized end sits GNOS-PV01, from Geneos Therapeutics: a DNA vaccine encoding up to 40 somatic neoantigens identified from each patient’s biopsy. Its Phase 1 trial (NCT04015700, the GT-20 study), run at Washington University’s Siteman Cancer Center, vaccinated nine patients with MGMT-unmethylated glioblastoma after surgery and radiation. The results, published in Nature Cancer, show a median progression-free survival of 8.5 months, a median overall survival of 16.3 months, and 24-month survival of 33%, including one survivor alive past 48 months. That’s a real result, from an open-label, single-arm, uncontrolled trial in nine patients: encouraging as proof of concept, but exploratory by design — and carrying the structural cost we already know: six to twelve weeks of individualized manufacturing per patient.

At the shared-antigen end, CVGBM, CureVac’s lipid-nanoparticle mRNA vaccine encoding eight tumor-associated-antigen epitopes, five HLA-A*02:01-restricted class I epitopes and three class II, may offer the sharpest lesson for any off-the-shelf platform, EVX-05 included. When direct immunopeptidomics was run on human cell lines, it confirmed presentation of most of the class I epitopes; none of the three class II epitopes in the design were detected. Computationally predicting that an antigen is “shared” and “conserved” doesn’t, on its own, guarantee it gets processed and presented physically and in CVGBM, that gap was asymmetric across HLA classes, not uniform.

UCPVax, for its part, bets on a single universal antigen (human telomerase) and has shown, in earlier trials in other indications, durable CD4+ immunogenicity with epitope spreading. Precision matters here: its glioblastoma-specific trial, MATVAC-1 (NCT07347210), is designed for 98 participants and, per the most recent public listing, hadn’t started recruiting yet. The immunogenicity data cited belongs to earlier UCPVax trials in other indications, not to this trial, which hasn’t produced its own results yet.

The blind spot shared by nearly every shared-antigen glioblastoma vaccine:

Computationally predicting that an antigen is “shared” and “conserved” doesn’t guarantee it will be processed and presented on the tumor cell surface. In CVGBM, the most advanced off-the-shelf candidate to test this empirically, direct immunopeptidomics showed a marked gap between what was predicted and what was actually presented, especially among the class II epitopes. EVX-05 hasn’t reached that test yet.

The most instructive precedent, though, isn’t biological, it’s institutional. Gritstone Bio built its GRANITE platform on a notably sophisticated AI-driven neoantigen-prioritization engine, EDGE™, aimed at predicting presentation and immunogenicity, and reached Phase 2 in microsatellite-stable colorectal cancer, with interim data showing a favorable trend in progression-free survival that never reached solid statistical significance. On October 10, 2024, the company filed a voluntary Chapter 11 petition in Delaware bankruptcy court. Gritstone’s outcome illustrates what can happen when a scientifically sophisticated platform needs more time and capital than a public company’s financial clock allows. It doesn’t prove that mismatch was the sole cause of the restructuring, the company itself also cited a delayed COVID-19 trial and an earlier workforce reduction, but it does illustrate the tension.

Two Clocks: Biological Validation and Corporate Necessity

What connects EVX-05 to Gritstone and to CVGBM isn’t negligence or cynical marketing, reading it that way would be a mistake, because it wouldn’t explain the pattern. It’s what complex-systems engineering has been describing since the 1980s as fragility relocation: adding a layer of technological sophistication to a system rarely eliminates the underlying risk. More often, it shifts it to another point in the chain usually a less visible one.

In EVX-03’s case, the risk the company most likely needed to solve was financial and logistical: individualized plasmid DNA manufacturing is expensive and slow, and that cost squeezes the cash runway of any public company reporting quarterly. If that reading holds, Evaxion didn’t eliminate that risk by discontinuing the program, it relocated it. Part of the clinical execution shifts outside Evaxion’s direct operating structure, toward an academic collaboration whose split of responsibilities and costs isn’t public. Part of it possibly shifts toward the pace of market communication, by pairing a discontinuation with a launch in the same filing. And part of it, with certainty, shifts toward the future: any failure in the remaining evidence rungs, actual presentation, in vitro immunogenicity, in vivo efficacy, preclinical safety, will surface after the announcement, not before.

Compared with Gritstone, which hit that same tension already at Phase 2, with substantially more capital committed, Evaxion’s sequencing is, strictly speaking, more prudent: it relocated the risk at the preclinical stage, where the cost of doing so is still low. But prudent isn’t the same as validated, and both clocks (biological and financial) will keep running in parallel until one catches up with the other.

What Evidence Would Actually Change Our Confidence

None of this means Evaxion’s chosen sequence is wrong, or that the dark-genome hypothesis is weak. Nor does it mean a public company, legally obligated to disclose material portfolio changes, should wait to complete all biological validation before announcing a program, that wouldn’t be realistic or fair to demand. The problem isn’t announcing a preclinical-stage candidate. It’s that the language of the announcement can convey more confidence than the available evidence supports.

The strongest constructive proposal, then, isn’t “announce later.” It’s announce precisely: present the program, show explicitly which rung of the evidence ladder it occupies, clearly separate computational discovery from confirmed experimental validation, and publish in advance the criteria that would advance the program (or shelve it) at each subsequent stage. That would make the next EVX-05 announcement, whenever it comes, verifiable on its own terms, not just favorable on the terms of whoever writes it.

That kind of transparency carries a real exposure cost: it’s easier to communicate enthusiasm than calibrated uncertainty. But it’s exactly the kind of trade-off (portfolio enthusiasm versus evidentiary precision) that any clinical-development or investment committee should be able to name explicitly, rather than infer between the lines of a Form 6-K.

Closing: Back to the Microscope

I’ll close by going back to that slide from my first year of residency. What Dr. Casavilca pointed out under the microscope, the necrosis, the hemorrhage, the heterogeneity, is exactly what any therapeutic glioblastoma vaccine has to overcome, and exactly what no computational data, however elegant the hypothesis behind it, can overcome on its own. Between that slide and the press release announcing a vaccine candidate lies a distance that only biological evidence (not the corporate calendar, not the technical vocabulary) can close.

The dark genome may well be one of the most promising answers immuno-oncology has for tumors like this one. But the question actually worth mapping isn’t whether the hypothesis is good. It’s what would have to change so that the language of an announcement accurately reflects a candidate’s real position in the evidence chain, so that, next time, a reader could tell, just from reading the release, exactly how much has been demonstrated and how much is still being predicted.

References:

Evaxion A/S. AACR Annual Meeting poster presentation, presented April 22, 2026; data released April 17, 2026. Available via: SEC Form 6-K.
Evaxion Biotech A/S. Form 20-F [Internet]. Washington: U.S. Securities and Exchange Commission; 2021-2024.
Evaxion A/S. Form 6-K [Internet]. Washington: U.S. Securities and Exchange Commission; Aug 17, 2026. Available at: sec.gov/Archives/edgar/data/1828253/000117184326005577/exh_991.htm [confirm the filing contains exactly the claims attributed here regarding portfolio management and Duke’s involvement].
Evaxion A/S. Business update and second quarter 2026 financial results [Internet]. Aug 20, 2026.
Garfinkle EAR, Perales-Linares R, Gimple RC, et al. Adjuvant personalized multivalent neoantigen DNA vaccination for MGMT unmethylated glioblastoma: a phase 1 trial. Nat Cancer. 2026 May 12. DOI: 10.1038/s43018-026-01163-w. ClinicalTrials.gov: NCT04015700.
Pre-clinical development of CVGBM: A therapeutic mRNA-based multiepitope vaccine for glioblastoma [abstract 22P]. Ann Oncol. 2024. Available at: annalsofoncology.org/article/S0923-7534(24)01550-3 [complete author list and volume/page data before publication]. Associated trial: ClinicalTrials.gov NCT05938387 (CV-GBLM-001).
ClinicalTrials.gov [Internet]. NCT07347210: Evaluation of UCPVax Vaccine +/- Pembrolizumab Combined With Standard Treatment (MATVAC-1). Bethesda: National Library of Medicine.
Gritstone bio, Inc. Gritstone bio Takes Action to Preserve Value and Strengthen Capital Structure; Voluntarily Files for Chapter 11 Restructuring [Internet]. Form 8-K, Exhibit 99.1. Washington: U.S. Securities and Exchange Commission; Oct 10, 2024. Available at: sec.gov/Archives/edgar/data/1656634/000119312524235637/d119736dex991.htm
Pitter KL, Tamagno I, Alikhanyan K, et al. Corticosteroids compromise survival in glioblastoma. Brain. 2016;139(Pt 5):1458-1471. [verify exact citation before publication]
Kassiotis G, Stoye JP. Immune responses to endogenous retroelements: taking the bad with the good. Nat Rev Immunol. 2016;16(4):207-219. [verify exact citation before publication]
Louis DN, Perry A, Wesseling P, et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro-Oncology. 2021;23(8):1231-1251. DOI: 10.1093/neuonc/noab106.
Quail DF, Joyce JA. The Microenvironmental Landscape of Brain Tumors. Cancer Cell. 2017;31(3):326-341. [verify exact citation before publication]

Confirming Isn't Curing: What a canceled oncology program, an AI CEO, and a student on GPT-4 have in common

Dr. Luis Cano — Fri, 21 Aug 2026 09:28:06 GMT

A few days ago, August 16, 2026, to be exact, I came across this tweet.

My first reaction was dismissive: another AI enthusiast telling the world robots will take over and none of us will have to work again. Then I looked closer and saw who had written it. That’s when everything changed.

The tweet was written by Dario Amodei, CEO of Anthropic, currently the most important AI company in the world.

Let’s be serious for a moment. Curing most diseases in five to ten years is a claim that lands differently depending on who you are. It excites a tech optimist. It gives hope to a patient. It triggers skepticism in a physician. And if you’re a pathologist, it makes you want to open the body and look.

The claim isn’t empty hype. Amodei is not a huckster, and his argument leans on things that already happened. AlphaFold solved, in months, a problem (predicting how a protein folds from its sequence) that had resisted decades of effort. mRNA vaccines, which used to take years to design, were adapted to a new virus in weeks. CAR-T therapy already produces durable, functionally curative remissions in certain leukemias. If AI multiplies the speed at which these breakthroughs occur, five to ten years stops sounding like fantasy and starts sounding like optimistic arithmetic.

That’s when the pathologist in me switches on. I reach for the dissection kit and start the autopsy, because underneath that arithmetic sits a formula almost nobody says out loud: disease → molecular mechanism → drug → cure. Four steps, one after another, as if curing a person were simply a matter of finding the right arrow between step two and step three.

It’s a crude simplification, but it’s the one we use. So before going further, some context is in order, otherwise everything that follows might sound like technology is the villain, and that isn’t accurate. AI did not invent this equation. It inherited it from institutional medicine, which (long before any language model existed) already privileged what is measurable, standardizable and easily attributable over what also determines whether a person is doing better, but costs more to measure. The danger isn’t that AI makes this mistake. It’s that AI can now execute it at a scale, speed, and authority that didn’t exist before.

How a Person Becomes a Variable

No one consciously decides to ignore the human being. Institutional reductionism isn’t born from bad intent or intellectual limitation. It’s born from an incentive architecture that rewards, at every link in the chain, whatever can be isolated, measured, and attributed. A clinical trial needs comparable endpoints across arms. A regulator needs to attribute an effect to a specific intervention, not to the diffuse sum of a patient’s life. A pharmaceutical company needs to isolate the value of its product to justify developing it. An AI model needs to turn success into an optimizable objective function. An insurer needs reproducible criteria for what it reimburses. A scientific journal needs variables another lab can replicate.

No single actor in this chain has to decide “let’s ignore the patient’s context.” Each one just has to reasonably optimize the piece of the problem in front of them. Reductionism emerges from the sum of those local optimizations, not from one decision someone could have made differently.

And precision matters here, because a clinical reader would rightly push back: medicine has developed real instruments for measuring quality of life, function, and patient experience, patient-reported outcomes, quality-of-life scales, shared decision-making, palliative care, functional assessments. The problem isn’t that these dimensions are entirely absent from the system. It’s that they routinely occupy a secondary position relative to endpoints that are easier to standardize, attribute, and convert into a regulatory or reimbursement decision.

To see what gets left out even when the system works exactly as designed, consider a case: a thirty-year-old patient with obesity, hypertension, chronic joint pain, and untreated depression receives an antihypertensive. Within weeks, blood pressure normalizes. The intervention worked: the biomarker dropped, immediate cardiovascular risk fell. But what, exactly, got cured?

The metabolic architecture that produced that risk in the first place is still intact. The joint pain keeps this person from the exercise that would sustain the improvement long-term. Untreated depression undermines adherence to the rest of the treatment. If the environment doesn’t provide easy access to food that supports real metabolic change, the biomarker improvement is being propped up by drugs, not by recovered health. And maybe (this is what no endpoint ever asks) what this person actually values isn’t a lower number on a lab report, but walking without pain again. None of this invalidates the antihypertensive. It just shows that its efficacy occupies one layer of the problem, not the entire system that determines whether someone is, or isn’t, better.

When the Variable Works, and When It Stops Being Enough

It would be dishonest to present the four-step equation as wrong in every case. There are diseases where it works extraordinarily well. Direct-acting antivirals cure hepatitis C in the vast majority of patients who complete treatment, though the real population-level outcome still depends on the infection being diagnosed, on access to the drug, and on the absence of reinfection, which are exactly the contextual variables this essay keeps pointing to. A well-chosen antibiotic resolves bacterial pneumonia by attacking its cause, though the real outcome of that pneumonia also depends on the specific pathogen, prior resistance, the patient’s age, and how quickly they received care.

You don’t need to claim total independence from context to hold the distinction. It’s enough to say this: in certain diseases, one dominant cause explains a large enough share of the problem that a molecular intervention produces something very close to a cure. In others (the obesity-with-comorbidities case above, much of solid-tumor oncology) no dominant cause exists, and the four-step equation falls short no matter how well it’s executed.

Lu-NeoB is the case that shows that limit best. It’s one of four discontinued oncology programs I autopsied a few weeks ago, alongside tilatamig samrotecan, AZD4360, and AZD2068, and the only one that matters for this argument, because it didn’t fail for lack of sophistication. It used the most advanced patient-selection tool available for that program: a paired theranostic PET scan with ⁶⁸Ga-NeoB, capable of showing in real time whether each patient’s tumor expressed the molecular target before treatment was ever administered.

The best selection tool available confirmed the target’s presence. That presence still wasn’t enough to produce the clinical outcome the program needed.

I’m not bringing this up as a total refutation of the equation, Lu-NeoB followed it with an exceptional degree of sophistication, but as a demonstration of its limits. The imaging could confirm the molecular target was present in the tumor. Confirming presence is not the same as guaranteeing that the rest of the tumor’s and the patient’s biology would behave the way the model needed in order for that presence to translate into clinical benefit. If molecular reductionism can fall short even when executed with the highest precision available, the promise of curing most disease by accelerating molecular discovery isn’t just optimistic, it’s a civilizational-scale version of the same experiment we already watched fall short at the scale of a single clinical program.

We’ve Seen This Movie Before

The pattern didn’t start with AI. On June 26, 2000, in the White House, Bill Clinton announced the first draft of the human genome with a line that, twenty-six years later, sounds almost lifted from Amodei’s: he said genome science would revolutionize the diagnosis, prevention, and treatment of most, if not all, human diseases. Francis Collins, who led the project, described that moment as the start of a new era in which we would understand the biological basis of disease more deeply and develop more effective ways to diagnose, treat, and prevent it.

Ten years later, Collins himself wrote an article in Nature: “Has the revolution arrived?”, honestly acknowledging the limits of clinical translation. Genome sequencing transformed basic research, the diagnosis of rare monogenic diseases, and molecular characterization in oncology. But curing or preventing most common diseases didn’t happen on the timeline that day’s euphoria at the White House implied. The reason wasn’t that the genome was useless. It was that most prevalent diseases aren’t driven by a single mutation, but by a polygenic architecture that interacts with environment, diet, and behavior, exactly the variables the four-step equation doesn’t capture.

A decade earlier, in 1990, another promised revolution had begun: gene therapy. A team led by W. French Anderson treated two girls with severe immunodeficiency by inserting a functional copy of the gene they were missing directly into their cells. At the time, it was global news, covered by the press as the start of an era of one-dose genetic treatments. But in 1999, Jesse Gelsinger, a patient in a different gene-therapy trial, died from a massive inflammatory response to a viral vector. In 2002, several children treated for severe immunodeficiency in France developed leukemia because of how the vector had inserted itself into their DNA. Regulators suspended multiple trials. Gene therapy wasn’t a false idea (approved, effective gene therapies exist today for specific indications) but it took nearly thirty years to become a clinical reality, and it still faces significant cost, infrastructure, and eligibility barriers that limit how accessible it actually is.

Major biomedical revolutions rarely turn out to be false. They turn out to be incomplete in ways we only discover once they start working.

AI as an Epistemic Amplifier

AI isn’t obligated to repeat the four-step equation. In principle, it can integrate a patient’s full clinical history, imaging, language, social variables, and longitudinal data. Reductionism doesn’t come from a limitation in the architecture. It comes from what we choose to feed it, and that happens in four moves worth separating.

First, the data: a model can only represent what was recorded; whatever was never documented doesn’t exist for it, even if it was decisive for the patient’s outcome.
Second, the objective: the model doesn’t decide what success means, someone converts an institutional priority into an optimizable metric before the model ever starts training.
Third, the proxy: since a person’s full health is nearly impossible to optimize directly, we use substitutes (tumor shrinkage, blood pressure, hospital length of stay) that capture part of the outcome, never all of it.
Fourth, authority: the more sophisticated the model appears, the easier it becomes to forget that its answer is only valid within the initial, partial, definition of the problem.

The point isn’t that AI simplifies because it’s intellectually limited. It’s almost the opposite: its ability to brilliantly solve a partial problem can hide the fact that the problem was defined partially in the first place.

Reductionism doesn’t start when we eliminate variables. It starts when a representation works so well we forget it was ever a representation. A biomarker partially represents a person’s health. An AI model partially represents the patient it describes. An AI assistant’s answer partially represents the reasoning it would have taken to reach that answer on your own. In all three cases, the representation is genuinely useful. The danger shows up when that usefulness makes us forget the difference between the map and the territory.

There’s a metaphor that captures this better than I can state it directly, and it comes from a comic-book story most people know: Spiderman who bonds with an alien symbiote called Venom, gains real power almost immediately (more strength, more speed) but over time, the symbiote starts quietly reshaping his judgment, without him noticing. It doesn’t make him weaker. It makes him more efficient, and in his case, efficiency means finishing off the enemy as decisively as possible. I think something similar happens with AI. It amplifies our capabilities, and slowly, it changes what we consider sufficient and, more importantly, what we consider human.

But reality tends to outdo fiction. Bastani et al. (PNAS, 2025) measured something equivalent in nearly a thousand high school students solving math problems with access to GPT-4. With unrestricted access, performance improved 48% while students used the tool, but once it was taken away for the final exam, those same students scored 17% worse than students who’d never had access at all. A third group, given a version of the same AI designed to guide with questions instead of handing over the answer (a Socratic tutor, not an oracle) improved 127%, with no drop-off afterward. The difference wasn’t the technology. It was whether the design preserved the friction necessary for learning, or eliminated it entirely.

Bastani and Lu-NeoB don’t demonstrate the same phenomenon (one is about human learning, the other about tumor biology) but they share a structure worth naming: in both cases, an intermediate signal improves or is present (solving a problem correctly, confirming a molecular target) and resembles the final outcome closely enough that we forget it isn’t the final outcome yet.

Who Decides What Counts as Cured

There’s a trap that’s easy to fall into when trying to answer this kind of simplification: adding more columns to the problem. Stacking nutrition, psychology, environment, income, and adherence into the model, as if the whole person would show up once enough variables were in the equation. It improves the representation. But it still reduces the person to a sum of data points, just with more rows.

Because something is missing that no dataset captures on its own. Picture two patients with the same cancer diagnosis and the same statistical prognosis. One accepts severe toxicity for a small chance of living six more months. The other prioritizes lucidity, time at home, the chance to say goodbye. A model, however sophisticated, can calculate with precision which intervention maximizes expected survival. What it can’t or more precisely, shouldn’t, adjudicate on its own is which of those outcomes is worth maximizing. Not because it lacks the capacity to process preferences, but because those preferences aren’t an external fact for the system to discover. They’re values each person has the right to express, revise, and negotiate. The distinction matters: it isn’t that AI lacks the capacity to calculate. It’s that it lacks the legitimacy to decide what’s worth calculating in the first place.

The patient isn’t merely a source of context for optimizing a decision someone else makes on their behalf. They’re the one with the authority to decide what should be optimized in the first place.

I don’t know whether most human disease will be curable by 2036. Nobody knows for certain, not even the person projecting it. But once we finally have the answer, there’s a question that matters more to me than the date: what will we have had to restore in a person before we’re willing to say we’ve cured them? That friction we lose, in the clinic or in the prompt, is the same friction that for centuries forced us to reason before we acted. The question isn’t whether AI can cure disease faster than we do today. It’s whether, by the time we find out, we’ll still have the muscle to ask the question correctly.

References:

Amodei D. Machines of Loving Grace [Internet]. Oct 2024. Available at: https://darioamodei.com/essay/machines-of-loving-grace
Amodei D. [Reply on X regarding AI’s public messaging and the 5-10 year disease-cure projection] [Internet]. Aug 16, 2026. X (Twitter). @DarioAmodei.
Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: evidence from high school mathematics. Proc Natl Acad Sci USA. 2025;122(26):e2422633122. DOI: 10.1073/pnas.2422633122. PMID: 40560616.
Collins FS. Has the revolution arrived? Nature. 2010;464(7289):674-675. DOI: 10.1038/464674a. PMID: 20360716. PMC: PMC5101928.
Collins FS. Testimony before the Senate Committee on Health, Education, Labor and Pensions [Internet]. Jul 2001. Available at: https://www.genome.gov (Speeches & Testimony Archives).
Clinton W. Remarks by the President on the Completion of the First Survey of the Human Genome [Internet]. White House, Jun 26, 2000. Available at: https://clintonwhitehouse3.archives.gov/WH/EOP/OSTP/html/00628_2.html
Blaese RM, Culver KW, Miller AD, et al. T lymphocyte-directed gene therapy for ADA-SCID: initial trial results after 4 years. Science. 1995;270(5235):475-480. [Anderson/Blaese’s original 1990 trial].
Death of Jesse Gelsinger in a gene-therapy trial, University of Pennsylvania, 1999.
Leukemia cases from insertional mutagenesis in children treated for SCID-X1, France, 2002.

An Autopsy of Four Drugs: The Cause of Death for Four AstraZeneca and Novartis Oncology Programs

Dr. Luis Cano — Fri, 14 Aug 2026 09:58:55 GMT

A pattern has surfaced over the past few weeks. Not in digital pathology (that gets a rest this week) but in the industry itself. Scanning the news, two stories stood out: cancelled drugs, pipelines shut down in the name of executive productivity. Two giants, AstraZeneca and Novartis, said goodbye to promising therapies.

So I asked myself, the way any clinician would: is this a symptom, or the disease itself? I decided to run an autopsy (I can’t help being a pathologist) to find out what had killed these programs. Profitability matters, and I don’t say that lightly: developing a drug takes years, enormous resources, and hard calls about where to keep investing. Ignoring that would be naive. With that caveat in mind, let’s work through these programs the way I’d work through a difficult case: history, findings, and an autopsy that may not hand us a single cause of death.

What the Family Tells Us

July 27, 2026. Pascal Soriot, CEO of AstraZeneca, is on a quarterly earnings call, walking analysts through Q2. Someone asks about three programs that just vanished from the pipeline, two antibody-drug conjugates and one radioconjugate, all cancelled the same quarter. Two of the three, tilatamig samrotecan and AZD4360, were grown in-house. The third, AZD2068, arrived through the 2024 acquisition of Fusion Pharmaceuticals, but by then it was already running under the same pipeline discipline as the other two.

Soriot doesn’t apologize. He frames it as discipline.

“Very important... to fail fast in the early phase,” Soriot told Fierce Biotech on the call. “When we are positive, we move into phase quickly. But we want to fail fast so we can redeploy resources to other projects.”

A few lines down in the same release, Susan Galbraith (AstraZeneca’s oncology and hematology chief) is celebrating different news: sonesitatug vedotin, an ADC against Claudin 18.2 (CLDN18.2) licensed from biotech KYM, just hit its primary endpoint in a phase 3 gastric cancer trial.

The trial met dual endpoints for overall survival in third line and progression-free survival in CLDN18.2-positive gastric cancers. Galbraith described it as having the potential to “reshape the treatment of gastric cancer by replacing classic chemotherapy.”

One of the three programs just buried, AZD4360, targeted the exact same protein: Claudin 18.2. Same target. Same day. Opposite results.

Fail fast, Soriot says, is discipline. Before we go further, hold onto a question: does failing fast leave you time to understand why you failed, or does it just give you permission to stop looking?

A Simple Diagnosis That Seems Sufficient

We could stop right there. Almost no one questions that when someone doesn’t wake up the next morning, the most likely cause was a heart attack. That holds for certain patients. But when the clinical picture doesn’t fit the age group, something doesn’t add up, an unseen condition, a rare disease. A diagnosis that simple is never enough for a pathologist, you run the autopsy anyway.

AZD4360 enrolled only patients with documented CLDN18.2 expression, confirmed by immunohistochemistry on biopsy, the standard method, unchanged for twenty years: a tissue sample, a stain, a pathologist reading a positivity cutoff under a microscope. A single fixed snapshot of a tumor that keeps changing after that photo is taken.

Trial D8930C00001 (NCT06921928) evaluated AZD4360 in a study projected for up to 117 patients with pancreatic adenocarcinoma, gastric/gastroesophageal junction cancer, and biliary tract cancer, all with CLDN18.2 confirmed on baseline biopsy. AstraZeneca’s stated reason for discontinuation, verbatim: “due to efficacy.”

Tilatamig samrotecan (AZD9592), a bispecific EGFR/cMET ADC, followed the same selection logic, histological confirmation, with no strict quantitative cutoff for co-expression.

The EGRET trial (NCT05647122) was projected for up to 403 patients with advanced solid tumors (lung, head and neck, colorectal) as monotherapy and in combination with osimertinib, 5-fluorouracil, leucovorin, and bevacizumab. Discontinued for “unfavorable PK/PD.”

AZD2068, the Actinium-225 radioconjugate against the same target pair, met the same fate in a study projected for 110 patients, with the same official reason: PK/PD.

Behind those numbers, 403, 117, 110 patients projected, of which only tens are publicly known to have actually been treated, with no full breakdown by program, are real people. They enrolled in a phase 1 trial knowing, because informed consent told them so, that the odds of direct benefit were low. Picture what it meant for someone with biliary tract cancer, likely already through standard chemotherapy with no approved options left, to travel to a trial site, undergo an extra biopsy just to confirm CLDN18.2, and take an experimental drug, not expecting a cure, but because there was nothing else left to try. That calculation, repeated across every enrolled patient, is the human cost behind a word as clinical as “efficacy” in a press release.

My report so far could write its own conclusion: the selection method caused the heart attack. A static biopsy, a binary positivity cutoff, can’t capture something as dynamic as whether a drug will distribute well, stay in place long enough, release its payload where it should. Of course they failed. The tool is too primitive for the question it’s being asked to answer.

I could call it a day, everything documented, the case closed given the outcome. But I had four cases on my table, and so far I’ve only worked through three. Will the last one tell me the same story, letting me close the file? Or will it show me something else?

The Fourth Case: Lu-NeoB

If I left it to the math, I’d say the same thing happened here: similar patterns, same symptoms, same outcome. Medicine taught me never to close a diagnosis before examining the tissue. And this case taught me something more uncomfortable: sometimes even the microscope isn’t enough. Looking closely, something doesn’t fit. Lu-NeoB didn’t depend on a static biopsy or an arbitrary IHC cutoff. Novartis used something far more sophisticated.

The radioligand ¹⁷⁷Lu-NeoB, targeting the GRPR receptor, had its own theranostic pair: ⁶⁸Ga-NeoB, a diagnostic version of the same molecular family, labeled with Gallium-68 instead of Lutetium-177, used as a PET/CT tracer before treatment. Not a tissue snapshot, a functional, real-time confirmation in the living patient.

Eligibility in the NeoRay trial (NCT03872778) required at least one tumor lesion to take up ⁶⁸Ga-NeoB at an intensity equal to or greater than the spleen’s, with more than 50% of lesions visible on CT/MRI also PET-positive.

Seventeen patients went through the phase 1 dose-escalation cohort. Seventeen people with advanced solid tumors who, beyond the usual biopsy and consent, underwent an additional PET scan, diagnostic radiation stacked on top of the therapeutic dose to come, to prove, with a precision no AZD4360 patient ever had to prove, that the drug actually reached their tumors.

Phase 1 dosimetry results, published in the Journal of Nuclear Medicine, showed a favorable profile in healthy organs: 0.10 Gy/GBq in the kidneys, 0.055 Gy/GBq in the pancreas, 0.018 Gy/GBq in red bone marrow, well below radiological tolerance limits. Maximum Tolerated Dose was set at 9.25 GBq, with dose-limiting toxicities documented at 11.1 GBq.

The tracer confirmed uptake in at least part of the disease, while dosimetry showed acceptable exposure in the organs evaluated. Everything the prior diagnosis said should be missing from an ADC with binary IHC selection, Lu-NeoB had, at least as far as phase 1 dosimetry can show.

In July 2026, Novartis discontinued the program anyway.

The company’s official, verbatim reason: “early clinical data did not support advancing it to the next stage of development.” And, explicitly: “this decision was not based on any new or unexpected safety concern.”

This is where the earlier diagnosis breaks down. The static biopsy and the arbitrary cutoff no longer explained the failure. The break had to be somewhere further down the chain, even if public data can’t pinpoint where.

There is, however, a mechanistically plausible hypothesis, suggested by the physics of the isotopes themselves, that no PET image alone can resolve: capturing is not the same as retaining. Gallium-68 has a half-life of 68 minutes, an instant snapshot of binding affinity, taken minutes after injection. Lutetium-177, the isotope doing the therapeutic work, has a half-life of 6.7 days. If the Lutetium-labeled radioligand clears the tumor faster than its short-lived diagnostic twin predicted, the accumulated radiation dose may never reach the threshold needed to lethally break tumor DNA, no matter how positive the PET looked that morning. It’s a hypothesis consistent with what we know about the pharmacokinetics of both isotopes. It is not, yet, a confirmed explanation.

What public data can’t tell us is how long the therapeutic radioligand stayed in each lesion, or what dose each one actually absorbed. Confirming the target is there, even with functional imaging in a living patient, doesn’t by itself confirm the vehicle stays long enough to do the damage it promises.

Where the Chain Breaks

Looking at these four cases laid out in front of me, the full picture comes into view: therapeutic response isn’t a binary event that hinges on whether the target is present. It’s the output of a sequential chain, target presence, tumor tissue access, receptor retention, sustained payload delivery or radiation dose, subcellular biological damage, and only then, measurable clinical response. One broken link is enough to void everything upstream, no matter how solid.

In ADCs, that break tends to happen at the interface between binding and release: an affinity too high can make the drug pile up on peripheral tumor cells without penetrating the core; a linker too unstable releases the payload in the bloodstream too early; one too stable never releases it where it should. In radioligands, a possible break sits between uptake and retention, one of the hypotheses Lu-NeoB forces us to consider. Different mechanisms. What they share is that neither is visible to the diagnostic used to enroll the patient in the first place, whether that’s a twenty-year-old IHC test or a state-of-the-art functional PET scan.

It’s worth noting what neither company said, as much as what they did. Novartis confirmed Lu-NeoB wasn’t a safety issue but never clarified where in the chain it broke. AstraZeneca went a little further: it distinguished PK/PD from efficacy across its three programs, more than the industry offered a decade ago, when a discontinuation was usually summed up in a single word with no breakdown. But a two-word label still isn’t a complete, public mechanistic explanation.

Which brings back the question left hanging in the first section. Fail fast has a defensible logic: don’t keep funding something that doesn’t work. But the speed fail fast demands is, almost by definition, the same speed that makes it hard to get to the bottom of why something failed, not necessarily because no one investigates internally, but because that knowledge rarely gets shared or accumulated in a way that benefits the next program. This isn’t cynicism aimed at AstraZeneca or Novartis specifically, that would be unfair, and beside the point. It’s the same institutional blind spot as always, now applied to the management of failure itself: the system certifies quickly whatever it can certify quickly, because the time and cost of understanding and sharing the full chain don’t fit inside the cycle a quarterly report demands.

An Experimental Treatment as a Possible Solution

In 2021, the FDA’s Oncology Center of Excellence launched Project Optimus, and in 2023 published the formal guidance that gave it regulatory weight: an initiative to replace the classic dose-selection standard, Maximum Tolerated Dose, which essentially asks how much a patient can withstand — with Optimal Biological Dose, which asks something different: how much does the tumor need, measured dynamically, to trigger the biological modulation you’re after.

The FDA states explicitly that for targeted therapies like ADCs, bispecifics, and radioligands, efficacy doesn’t scale linearly with dose, and requires sponsors to justify dosing through randomized dose-response studies and early-stage pharmacometric modeling, not just confirmation that the antigen is present.

This isn’t a finished cure. At best, it’s the first stretch of a treatment we don’t yet know is viable at the scale and cost the industry needs. But it’s a verifiable regulatory fact, not wishful thinking, and there’s a second, more modest but real signal: AstraZeneca’s own ability to distinguish PK/PD from efficacy this quarter is evidence the diagnosis is slowly getting less crude than it was a decade ago, when the industry as a whole reported an attrition rate above 70% between phase 1 and approval for oncology antibodies, almost always with no breakdown by cause.

We haven’t arrived. But when these programs fail today, they fail leaving a slightly more legible trail behind.

The Warning Sign

I remember my emergency medicine attending’s voice, yelling at me while I worked on a patient in the trauma bay: “Details, Luis. Right now you’re just treating, but if you want to be effective, pay attention to the details.” That’s why Soriot’s phrase: fail fast, stuck with me. Because failing fast: does it let you understand why you failed, or does it just let you stop looking?

After these four cases, the answer isn’t a yes or a no, it’s a distinction. Failing fast without preserving and sharing what was learned can turn financial discipline into scientific amnesia. Failing fast alongside something like Project Optimus pushing from the regulator, with companies starting to publicly distinguish where the thread broke, could, over time, become something else: accumulated learning instead of quarterly amnesia.

As a pathologist, I’ve gone as far as the findings allow: I can describe the chain and point to where it may have broken, but I can’t certify a cause of death that public data doesn’t reveal. Still, there’s a feeling I can’t shake, the choice between failing fast and moving on, or failing and dissecting the error to learn from it. Today, the industry has both routes drawn out, and the difference between them won’t be decided by a more precise biomarker. It will be decided by whether someone is willing to pay the slow, unphotogenic cost, in a quarterly report, of understanding the full chain instead of just labeling where it broke.

References

Waldron J. AstraZeneca execs ‘very confident’ in homegrown ADC pipeline despite scrapping 2 candidates. Fierce Biotech [Internet]. 2026 Jul 27 [cited 2026 Aug 13]. Available from: https://www.fiercebiotech.com/biotech/astrazeneca-execs-very-confident-homegrown-adcs-despite-scrapping-2-candidates
Novartis scraps mid-stage radiopharma asset as Pluvicto soars. BioSpace [Internet]. 2026 Jul 21 [cited 2026 Aug 13]. Available from: https://www.biospace.com/business/novartis-scraps-mid-stage-radiopharma-asset-as-pluvicto-soars
Novartis drops Phase 2-stage radioligand therapy, but CEO’s commitment to modality ‘unchanged’. Fierce Biotech [Internet]. 2026 [cited 2026 Aug 13]. Available from: https://www.fiercebiotech.com/biotech/novartis-drops-phase-2-stage-radioligand-therapy-commitment-modality-unchanged
Mittra ES, Solnes LB, van der Veldt AAM, Wong J, Aloj L, Heinrich MC, et al. Phase 1 study of [177Lu]Lu-NeoB in patients with advanced solid tumors overexpressing gastrin-releasing peptide receptor: preliminary safety and dosimetry results. J Nucl Med. 2026 Feb. DOI: 10.2967/jnumed.125.270411.
ClinicalTrials.gov. First in Human Study of AZD9592 in Solid Tumors (EGRET). NCT05647122 [Internet]. Available from: https://clinicaltrials.gov/study/NCT05647122
ClinicalTrials.gov. Study for AZD4360 in Participants With Advanced Solid Tumours. NCT06921928 [Internet]. Available from: https://clinicaltrials.gov/study/NCT06921928
ClinicalTrials.gov. A Phase 1, Dose-escalation Study of [225Ac]-FPI-2068 in Adult Patients With Advanced Solid Tumours. NCT06147037 [Internet]. Available from: https://clinicaltrials.gov/study/NCT06147037
ClinicalTrials.gov. [177Lu]-NeoB in Patients With Advanced Solid Tumors and With [68Ga] (NeoRay). NCT03872778 [Internet]. Available from: https://clinicaltrials.gov/study/NCT03872778
U.S. Food and Drug Administration. Project Optimus. Oncology Center of Excellence [Internet]. [cited 2026 Aug 13]. Available from: https://www.fda.gov/about-fda/oncology-center-excellence/project-optimus
Navigating the clinical progress of antibody-drug conjugates: Emerging opportunities and remaining challenges. Cell. 2026 May. Available from: https://www.sciencedirect.com/science/article/abs/pii/S0092867426004514

When Target Expression Isn't Enough

Dr. Luis Cano — Fri, 07 Aug 2026 09:58:04 GMT

THE SCENE

Every pathologist has seen a case like this. HER2-positive breast cancer, IHC 3+, the kind of staining that makes a slide worth showing to residents. The pathologist orders in situ hybridization anyway, just to document it properly. Some cases are too clean not to study closely.

In the mid-2010s, a phase II trial enrolled 164 patients with exactly that profile (HER2-positive by immunohistochemistry) to test, experimentally, trastuzumab emtansine combined with pertuzumab as neoadjuvant treatment, before surgeons decided how much tissue to remove. It was not yet standard of care. It was the question the trial existed to answer. Picture two of those patients, enrolled with the same pathology report in hand.

Neither knows yet that the outcome will diverge. Same waiting room, different dates, same question, will this work?, and the same reassuring answer: the tumor expresses the target, the drug was built for exactly this, there’s real reason for optimism.

Their oncologist has no reason to treat them differently. The pathology report says the same thing for both: “HER2 positive, IHC 3+.” Same box checked, same trial protocol. Both walk through the same door.

Months later, one has no residual tumor in the resected tissue, pathologic complete response, the best possible outcome in this setting. The other has substantial residual disease. Same drug. Same dose. Same molecular target, on paper.

What the pathology report didn’t say (because the protocol never asked) is that in one patient HER2 was distributed uniformly across the tumor, and in the other there were entire regions where amplification simply wasn’t present, subclones nobody looked for because nobody had to. The same binary label, “positive,” concealed two completely different tumor architectures.

THE SCENE HAS A REAL ANCHOR, METZGER ET AL., PHASE II NEOADJUVANT TRIAL

55% vs. 0%

Pathologic complete response rate in homogeneous vs. heterogeneous HER2 tumors (n=157 evaluable of 164 enrolled, 16 with intratumoral heterogeneity, p<0.0001, adjusted for hormone receptor status).

No hypothetical patients were needed to make this scenario real: in that trial, none of the 16 patients with intratumoral HER2 heterogeneity achieved a complete response. None. While 55% of those with a homogeneous tumor did. The pathology report, in both groups, said exactly the same thing: HER2 positive.

The word heterogeneity is worth pausing on, because in everyday clinical language it sounds like a footnote. In this trial it had a precise definition: a tumor region with HER2 amplification in fewer than 50% of cells, or an outright HER2-negative area, detected by FISH in biopsies from two different sites of the same tumor. A single biopsy wasn’t enough. The tumor had to be sampled in more than one place for the pattern to show up.

And there, perhaps, is the first hint of where this essay is headed: the problem isn’t only what we measure. It’s how many times, and in how many places, we bother to measure it.

THE QUESTION THAT SHOULDN’T MAKE SENSE

There’s something unsettling about a finding like this, and it isn’t only clinical. It’s epistemological. We’ve spent two decades building precision oncology on a simple premise: find the right molecular target, direct the drug at it, and the outcome should be predictable. HER2 positive. Trastuzumab. Benefit. That causal chain is, in large part, the founding promise of targeted oncology.

But the pathology report (the document that triggers that chain) was never built to capture whether the target was evenly distributed or patchy, whether the tumor had enough stroma to physically block drug delivery, whether the immune microenvironment was hospitable or hostile. The report answers a binary question: is the target there or not?, about a phenomenon that was never binary.

I’ve spent months noticing something similar in a field that, on the surface, has nothing to do with this: the race to build foundation models for digital pathology. The more I look at these two territories side by side, the more the shape of the problem looks the same, even though the content is entirely different.

This is the question that has followed me through most of what I write here: why do innovations that look inevitable take so long to change medicine. Almost every time I dig into that question, the answer isn’t the technology (which usually already exists, somewhere, in some paper) it’s the institutional architecture that decides which question gets solved first.

TWO OBJECTS THAT SHARE NOTHING

A foundation model trained on millions of histology images and a single-marker immunohistochemistry assay share no technology, no budget, no invention date. One is the leading edge of applied AI in medicine. The other is a staining method that has existed, essentially, for decades.

Foundation models chase perception without an equivalent decision infrastructure. Companion diagnostics chase reproducibility by reducing biology to whatever can be standardized. From opposite directions, both reveal the same gap: we can measure more than our institutions know how to use.

The foundation model industry (I’ve written about this before) has bet heavily on perception: models increasingly capable of “seeing” subtle tissue patterns, matching or exceeding the pathologist’s eye. What almost nobody is building with the same intensity is the decision infrastructure that would turn that perception into something clinically actionable, reproducible, auditable, transferable across labs and countries.

The companion diagnostic for an ADC runs the opposite path and arrives at the same place. It doesn’t chase perception (it’s deliberately simple, one marker, one binary cutoff), it chases reproducibility: the quality that lets it work the same way in any hospital, anywhere. The price of that reproducibility is reducing tumor biology to the one variable that’s easy enough to measure that way. Two opposite paths, perception without decision, reproducibility without full biology, and the same result: we measure more than our institutions know how to use.

That’s the blind spot these two objects, so different on the surface, share without knowing it.

It isn’t a coincidence that diagnosis and treatment selection took separate paths in recent precision oncology history. Diagnosis became AI’s territory, with research budgets in the hundreds of millions. Treatment selection for ADCs, meanwhile, stayed anchored to a last-century instrument (immunohistochemistry) because there was never an institutional moment where anyone asked whether both problems were, underneath, the same question asked twice.

WHY TISSUE MATTERS MORE THAN THE REPORT ADMITS

For an antibody-drug conjugate to work, the target has to be there, but that’s not enough. The drug has to physically reach it, crossing a tissue architecture the pathology report never describes in that level of detail.

Start with the intuitive part: tissue as physical obstacle. Dense extracellular matrix and proliferating cancer-associated fibroblasts act as a real barrier against diffusion of antibody-sized molecules. In some tumors, specific fibroblast subpopulations drive immune exclusion and physically constrict vasculature, restricting drug passage from blood vessel to tumor core. Chaotic tumor vascularization also creates perfusion gradients: central regions literally receive less drug than the periphery.

So far, intuition holds: more barrier, less drug. But biology, again, resists simplicity.

In third-generation conjugates like trastuzumab deruxtecan, cathepsin L, a protease overexpressed and secreted into the tumor microenvironment by stromal and inflammatory cells, can cleave the drug’s linker outside the cell, releasing payload directly into the interstitial space. This kills neighboring tumor cells that don’t even express the target (the so-called bystander effect) without requiring receptor-mediated internalization. It’s a finding demonstrated in specific models and still recent; there isn’t enough evidence to assume this mechanism operates with equal force in every patient or every cleavable-linker ADC, but it points at something real: the same stroma that acts as a barrier in one context acts as an activation mechanism in another.

Target heterogeneity, by contrast, doesn’t carry that same duality, it’s close to uniformly bad news, as far as the evidence shows.

There’s a third variable, less discussed than stroma or vasculature, that also decides the drug’s fate: the target’s subcellular location. HER2 retained at the cell membrane, where the antibody can bind easily, is not the same as HER2 displaced into the cytoplasm, partially out of reach. That distinction, membrane versus cytoplasm, is what the Normalized Membrane Ratio quantifies, a metric that resurfaces, with a starring role, in the final act of this essay.

None of these three variables (stroma, vasculature, subcellular localization) appears on a standard pathology report today. Not because they’re invisible to a trained eye: an experienced pathologist can often describe them qualitatively in the margin of a report. They’re absent because there’s still no quantitative vocabulary, reproducible across observers, that turns them into data a clinical trial can use to stratify patients. That, in essence, is the standardization problem running through this entire essay.

THE QUANTITATIVE PAYOFF, NOW IN ITS PLACE

55% vs. 0%

Metzger et al., spatial HER2 heterogeneity, not just its presence, determined whether pathologic complete response occurred. A factor invisible to the standard IHC assay.

WHAT THE COMPANION DIAGNOSTIC CHOSE NOT TO SEE

The regulatory design of ADC companion diagnostics isn’t an accident or a negligence, it’s a reasonable answer to a different problem than the one we now know exists. When these assays were designed, regulators needed to answer a simple, urgent question: does it make sense to expose this patient to an expensive, non-trivial-toxicity drug? A cheap, cross-lab-reproducible IHC assay, backed by decades of infrastructure already installed in hospitals worldwide, answered that question reasonably well.

What that design never incorporated is whether, within the “positive” group, subgroups exist with tissue architectures different enough to change their odds of benefit. Sacituzumab govitecan, a TROP2-directed ADC, doesn’t even require quantifying the target in its approved indication: the trial showed clinical benefit across such a wide expression range that a cutoff would have improperly excluded patients who did benefit.

No clinical guideline, not NCCN, not ESMO, requires or even suggests characterizing stromal, vascular, or immune infiltration variables to decide whether to prescribe an ADC as monotherapy today. Clinical practice still treats broad populations based almost exclusively on the binary presence or absence of the antigen. Not because the biological evidence doesn’t exist (as Act IV showed, it does, and it’s solid) but because there’s still no regulatory vehicle, no commercial incentive, built to carry it through to the clinical decision.

Regulatory review documents justify single-marker assays on concrete grounds: analytical simplicity, high baseline antigen prevalence in the target population, and the risk that too strict a cutoff would improperly exclude patients who’d still benefit clinically. It isn’t an arbitrary choice.

The practical result of that logic is a single-speed regulatory architecture: every new ADC reaches market with, at best, an expression assay for its own target, designed and validated in isolation. There’s still no framework to evaluate, across different ADCs, different targets, different payloads, whether the same set of microenvironment variables consistently predicts response. Every drug reinvents its own diagnostic question from zero. It’s the same pattern from Act III, again: an industry capable of investing in perception hasn’t yet found the institutional vehicle to invest with equal intensity in the decision infrastructure that would make it useful.

THE COUNTERARGUMENT THE THESIS ITSELF DEMANDS

It would be tidy to close here with a clean conclusion: the industry needs to reclassify patients by microenvironment, and digital pathology is the tool that will make it possible. That conclusion would be exactly the kind of overconfidence this essay has spent three acts criticizing elsewhere.

In the ASCENT trial, sacituzumab govitecan showed clinical benefit, in progression-free and overall survival, consistently across every TROP2 expression subgroup analyzed, including low expression. In TROPiCS-02, the finding was even more decisive: “there was no clear level of TROP2 expression at which a better treatment effect was observed,” as the trial’s principal investigator reported. Drug efficacy did not, in practice, depend on how much target was there.

THE EVIDENCE THAT LIMITS THE CENTRAL HYPOTHESIS

Benefit across every subgroup

ASCENT and TROPiCS-02: consistent clinical benefit from sacituzumab govitecan regardless of TROP2 expression level, including very low expression.

How do this finding and Metzger’s coexist in the same thesis? The answer, again, lies in drug architecture, not the microenvironment alone. Third-generation ADCs, cleavable linkers, permeable payloads, high drug-to-antibody ratio, appear potent enough to partially overcome both target heterogeneity and microenvironment barriers. The conjugate’s biological sophistication, in some cases, is already doing the work a microenvironment score (or a foundation model) would otherwise have to do from outside.

This is exactly what the Complexity-Value Matrix predicts: sophistication is a property of the problem, not automatically of the solution built on top of it. Sometimes the problem (tissue heterogeneity) demands an equally sophisticated answer, as with T-DM1. And sometimes the problem is already solved by drug design, and layering additional stratification on top adds no clinical value, only regulatory complexity and cost.

The question that actually matters, then, isn’t “do we need to measure the microenvironment?” in the abstract. It’s: for which ADC family, with which linker type and which payload, would that measurement change a real clinical decision? Most of the time, that question still doesn’t have an answer grounded in prospective evidence.

Put differently: HER2 heterogeneity matters enormously for T-DM1, because that drug depends almost entirely on receptor-mediated internalization inside the cell that expresses it, there’s no bystander effect to compensate for uneven tissue. TROP2 expression seems to matter less for sacituzumab govitecan: its pH-sensitive CL2A linker also enables continuous extracellular payload release in the acidic stroma. That doesn’t mean internalization stops mattering, available mechanistic evidence indicates both pathways contribute, it means the drug has a second mode of action that partially compensates when the first one fails.

But this shouldn’t be mistaken for a free solution. The same mechanism that lets an ADC bypass part of the tissue heterogeneity problem (a cleavable linker, capable of releasing payload outside the target cell) is, per available evidence, the one most associated with systemic toxicity. A meta-analysis of 40 clinical trials and roughly 7,900 patients found significantly more grade ≥3 adverse events in ADCs with cleavable linkers than non-cleavable ones, and higher drug-to-antibody ratio was independently associated with higher probability of severe toxicity.

THE HIDDEN PRICE OF THE “POTENCY WITHOUT CLASSIFICATION” PATH

47% vs. 34%

Grade ≥3 adverse events in cleavable-linker (bystander) ADCs vs. non-cleavable linkers, meta-analysis of 40 clinical trials (n=7,879). The same mechanism that compensates for tissue heterogeneity reduces drug selectivity throughout the body.

“Build a more potent drug” isn’t a neutral alternative to “classify the patient better.” It’s a different bet, with its own cost, except that cost is paid by the patient, in adverse events, not by the pharmaceutical company, in diagnostic research budget.

THE INCOMPLETE MAP

There is, however, one real exception, and it’s worth naming precisely, not inflating it. Based on public sources identified through August 2026, TROPION-Lung17 is the only phase III trial that prospectively randomizes patients based on a computational pathology parameter: TROP2’s Normalized Membrane Ratio, calculated through Roche and AstraZeneca’s Quantitative Continuous Scoring system, which received the FDA’s first Breakthrough Device Designation ever granted to an AI-driven companion diagnostic, in 2025. That designation opens a priority regulatory pathway, it does not, yet, equal demonstrated efficacy or definitive clinical validation.

THE ONLY KNOWN PROSPECTIVE EXCEPTION

1 of 1

TROPION-Lung17: the only identified phase III trial prospectively randomizing by a computational pathology biomarker. Ongoing; no published results to date.

It’s worth being precise about what that exception actually measures. NMR quantifies the ratio between membrane and cytoplasmic target, a subcellular distribution measurement, not a direct demonstration of how much payload ends up internalized in the cell. It showed predictive potential in retrospective analyses and is being validated prospectively; it does not measure stroma, vasculature, or immune infiltration. It’s a real advance in one of the four microenvironment domains biology tells us matter (the target itself) and it still has no published clinical results confirming it.

The map, then, is only beginning to be drawn, and only in one corner. The rest of the territory (stromal architecture, immune competence, vascular accessibility) still has no standardized measurement instrument, no companion diagnostic, no regulatory vehicle carrying it through to daily clinical decisions.

There’s an asymmetry here worth naming, not to judge it but to understand it. Six ADCs are approved today with biomarker selection, each with its own linker, its own payload, its own engineering optimized in isolation. On the other side, there’s a single computational pathology platform with regulatory designation, shared so far between two drugs in the same family. A more potent linker’s engineering doesn’t transfer to a competitor, it’s investment that doesn’t compound. A tissue classification infrastructure, once validated, does transfer from one ADC to the next. It’s worth asking what would happen if a fraction of what’s currently spent making each drug individually more potent and less selective were invested in that shared infrastructure instead, not because that’s, given today’s evidence, the correct answer, but because it’s the question current investment logic doesn’t seem to be asking.

And there’s a more uncomfortable question underneath that one, perhaps the real blind spot of this entire essay. Almost everything we’ve built in precision oncology, the IHC assay, the regulatory approval criterion, even the microenvironment score this essay imagines, is designed to produce a dichotomous answer about a phenomenon that was never dichotomous: expresses or doesn’t, responds or doesn’t, high score or low score. Tissue biology, as Act IV showed, is a continuum of grays, of barriers that are sometimes mechanism and sometimes obstacle, of heterogeneity that matters enormously in one drug and almost nothing in another.

Maybe the question isn’t how well we classify, or how potent we make the drug, but whether we’re building institutions capable of acting on an answer that was never yes or no, but a range of probability.

I don’t claim this essay closes that gap. What I have is the certainty of having seen the same shape of problem in two territories nobody puts side by side: foundation models chasing perception without decision infrastructure, companion diagnostics chasing reproducibility by reducing biology to what can be standardized. From opposite directions, both reveal the same gap, we can measure more than our institutions know how to use. I’m still drawing this map. If someone else is seeing the same pattern from another point in the territory, it probably draws better in company.

References:

Metzger Filho O, Viale G, Trippa L, et al. HER2 heterogeneity as a predictor of response to neoadjuvant T-DM1 plus pertuzumab: results from a prospective phase II clinical trial. J Clin Oncol. 2019;37(15_suppl):502.
Predictive Biomarkers of Antibody-Drug Conjugate Efficacy for Solid Tumors: Current Challenges and the Potential Role of Quantitative Proteomics. Clin Cancer Res. 2026;32(4):661.
Roche. Roche granted FDA Breakthrough Device Designation for first AI-driven companion diagnostic for non-small cell lung cancer [press release]. April 29, 2025.
Bardia A, Rugo HS, Tolaney SM, et al. Final results from the randomized phase III ASCENT clinical trial in metastatic triple-negative breast cancer and association of outcomes by HER2 and TROP2 expression. J Clin Oncol. 2024;42(15):1738-1744.
Rugo HS, Bardia A, Marmé F, et al. Sacituzumab govitecan (SG) vs treatment of physician’s choice (TPC): efficacy by Trop-2 expression in the TROPiCS-02 study. Cancer Res. 2023;83(SABCS22-GS1-11).
Daiichi Sankyo / AstraZeneca. TROPION-Lung17 TROP2 Biomarker Directed Phase 3 Trial of DATROWAY Initiated in Patients with Previously Treated Advanced Nonsquamous NSCLC [press release]. January 13, 2026.
Tang SC, Wynn C, Le T, McCandless M, Zhang Y, Patel R, Maihle N, Hillegass W. Influence of antibody-drug conjugate cleavability, drug-to-antibody ratio, and free payload concentration on systemic toxicities: a systematic review and meta-analysis. Cancer Metastasis Rev. 2024;44(1):18.

We Didn't Get Foundation Models Wrong. We Got the Question Wrong.

Dr. Luis Cano — Fri, 31 Jul 2026 09:57:06 GMT

In 1998, a group of GI pathologists ran an uncomfortable experiment. They took the same histological slides (esophageal, gastric, colorectal biopsies) and sent them to 31 expert pathologists in 12 countries. The question was simple: does everyone see the same thing?

They didn’t. On esophageal lesions, agreement between Western and Japanese pathologists dropped to 14%. Essentially chance. The same slide, under the same microscope, read as invasive carcinoma in Tokyo and high-grade dysplasia in Boston.

No one in that study failed to see what was on the tissue. The problem wasn’t anyone’s eyes. Each pathologist was carrying a different rule for deciding when an abnormal cell crosses the line into cancer. The Western school required demonstrated histological invasion of the lamina propria or submucosa. The Japanese school classified primarily by cytological atypia and nuclear hyperchromasia, without requiring proof of invasion. Two rigorously applied systems, producing incompatible labels on the same human tissue.

The consequence wasn’t academic. With disagreement at that scale, the clinical risk is obvious even without a single documented case: a patient diagnosed with carcinoma in Japan, read under Western criteria, would be classified differently, with the surgery already indicated at home now in question. Or the reverse: a malignancy call that, under the other system, would never have justified a total gastrectomy with lymphadenectomy. You don’t need a specific chart to see that two different labels on the same tissue, crossing a border, cost something.

This wasn’t solved by training anyone to “see better.” It was solved by negotiating an agreement: the 1998–2000 Vienna Classification, which restructured the categories into five uniform clinical-management tiers. Interobserver agreement rose from 14–37% to over 70%. No pathologist’s eye changed. The rule everyone agreed to follow did, and that distinction matters: Vienna didn’t erase the underlying disagreement between the two diagnostic schools. It translated both criteria into comparable management categories. The scientific disagreement is still documented in the literature. What changed is that it stopped automatically producing incompatible clinical decisions.

I open with this story because, at a smaller scale, it’s exactly the problem the medical AI industry keeps refusing to look at.

Subscribe now

The Infrastructure Nobody Asked For

Everyone is talking about foundation models. The dominant narrative, repeated in Nature Reviews Drug Discovery, Nature Medicine, and nearly every strategic conversation in the sector— is that these models are building digital pathology’s infrastructure layer, the way an operating system once structured personal computing. Since 2024, half a dozen of these models have shipped from academic and corporate labs, trained on banks of millions of whole-slide images:

Not all of these models pitch themselves as replacing the pathologist, many are, by their own technical description, general-purpose feature extractors, meant to power downstream applications someone else will build. But the ambition they share, explicitly or not, is that the core problem is perception: if a system learns to recognize morphological patterns more precisely, at greater scale, over more tissue, most of the clinical value follows from that. It’s a bet that performs well on academic benchmarks, and pulls in capital no standardization consortium ever has.

Worth naming a distinction here that carries the rest of this piece: computational complexity (processing gigapixel images, scaling transformer architectures, inferring genomic correlations from a routine stain) is not the same as institutional complexity getting two medical schools to agree on where a diagnostic line sits, and making that agreement auditable, defensible, and acceptable to a regulator. Foundation models are built to absorb the first. The question this piece is asking is what happens when the real problem is the second.

The largest independent benchmark run on these models to date (32 models, 41 tasks, over 17,500 slides, from Olivier Gevaert’s team at Stanford, in Nature Communications) offers evidence consistent with that reading: neither model size nor pretraining dataset size consistently predicts downstream performance. Differences among the top performers are, in most cases, statistically insignificant. And once models leave their training domain (from TCGA into real clinical data, a different hospital, a different scanner) the rankings shift. No single model dominates every setting. The same study found that ensembling several models does improve aggregate robustness, suggesting these models capture complementary strengths, even if none resolves the underlying problem alone.

In other words: scaling any one model isn’t buying the edge it promises. And that same study, by design, doesn’t evaluate (because no foundation model can, on its own) whether two hospitals in two countries will call the same thing carcinoma.

But there’s another bet. And someone placed it eight years before “foundation model” meant anything in medicine.

The Bet That Actually Worked: Instrument, Don’t Replace

In 2016, at the ISBI challenge on breast cancer metastasis detection in lymph nodes, the algorithm from Andy Beck and Aditya Khosla’s team (the founders of PathAI) scored an AUC of 0.925 on whole-slide classification. An expert pathologist, working alone with no time limit, scored 0.966. The algorithm, on its own, still lost to the human.

The finding that actually mattered was different: when the pathologist’s read was combined with the algorithm’s assistance, the combined AUC rose to 0.995, roughly an 85% reduction in the specialist’s error rate. This wasn’t a race between the human eye and the machine. It was a different question: what does the pathologist need in order to be wrong less often?

“A model can in no way replace the work of a pathologist. It should be used as a tool to help the pathologist identify problem areas within a tissue sample.”
Andy Beck, co-founder and CEO of PathAI

That early bet held. PathAI built AIM-NASH, the name under which the FDA and EMA qualified the tool as a Drug Development Tool, built specifically to measure and reduce interobserver variability in grading metabolic dysfunction-associated steatohepatitis within MASH clinical trials. The company was founded, per its own corporate materials, “to apply technology to pathology to reduce error rates and increase accuracy and reproducibility.”

Reproducibility. Not replacement. That was the bet, funded by a $60M Series B led by General Atlantic in 2019, at a moment when the rest of the sector’s dominant ambition was already something else. I don’t offer this as proof the whole industry should follow one strategy (it’s a case, not a law) but as evidence that the instrumentation-and-reproducibility bet is viable, gets funded, and has already cleared a regulator.

A Pattern We’ve Seen Before

The ambition of “the model just diagnoses” isn’t new, and it isn’t unique to pathology. It’s the same narrative pattern we’ve already lived through, with exact dates and names everyone recognizes.

On October 19, 2016, at an official press call, Elon Musk promised that by the end of 2017 a Tesla would drive fully autonomously from Los Angeles to New York, “without the need for a single touch” on the wheel. On April 22, 2019, at Tesla’s Autonomy Investor Day, he raised the stakes: “I feel very confident predicting autonomous robotaxis for Tesla next year… we’ll have more than a million on the road.”

Seven years since that first promise. We still have taxi drivers.

In November 2016, at a Creative Destruction Lab talk in Toronto, Geoffrey Hinton (today a Nobel laureate, then already deep learning’s most influential figure) said of radiologists: “People should stop training radiologists now. It’s just completely obvious that within five years deep learning is going to do better than radiologists.” Nearly ten years later, radiology is one of the highest-demand, best-paid specialties in the U.S., and of the more than one thousand FDA-cleared AI devices for radiology, not one is authorized to diagnose autonomously without a radiologist’s sign-off.

I’m not citing Musk or Hinton to cast them as villains. They’re just the most visible names in a much broader pattern: an entire industry that, faced with a hard problem, would rather promise to eliminate it than ask what’s causing it. “We’re going to replace the pathologist” is a headline. “We’re going to harmonize diagnostic criteria across twelve countries” is not. But only the second one actually fixes what costs the patient time, money, and safety.

Institutional Complexity, Not Computational Complexity

This is exactly what my Complexity-Value Matrix predicts: sophistication is a property of the problem, not of the tool used to solve it. I planted this distinction earlier, but it’s worth pausing on, because it’s this piece’s central conceptual contribution.

Computational complexity is processing gigapixels per image, scaling vision transformer architectures, inferring genomic correlations from a routine stain. It’s the complexity foundation models are built to absorb and where they genuinely excel.

Institutional complexity is something else entirely: getting two medical schools to agree on where displasia ends and carcinoma begins, folding that consensus into multidisciplinary tumor boards, making it auditable to a regulator, defensible in a lawsuit. It’s slow, it doesn’t scale with more training data, and there’s no benchmark you can win in a paper.

Applying massive computational complexity to a problem of institutional complexity is a category error. A feature extractor with billions of parameters doesn’t, by itself, resolve decades of disagreement between Japan and the West over the histological definition of a tumor margin. That problem isn’t won with scale. It’s won with governance.

This isn’t just my intuition anymore. The Stanford benchmark offers evidence consistent with this reading: comparing small, base, and large ViT architectures, and pretraining datasets from under 100,000 slides to over a million, scaling benefits are most consistent within TCGA, the same domain most of these models were trained on. On external and out-of-domain data, the benefits are weak, variable, and dependent on domain and task. Scale appears to buy more performance inside the distribution it trained on than outside it. It’s a correlation, not an airtight causal proof (the study itself flags this as a limitation) but it points the same direction as the Complexity-Value Matrix: computational complexity mostly solves computational problems. It doesn’t, by itself, get two hospitals to agree.

The Question Nobody Verified First

I don’t think we got foundation models wrong. I think we got it wrong expecting a better visual representation, on its own, to solve clinical reproducibility. These are related problems (better perception can help standardization, if it’s designed for that) but they’re not the same problem, and treating them as if they were is the choice worth revisiting.

Why did the industry bet so heavily on replacement instead of standardization? I can’t prove it was a purely commercial decision, that would claim more than the evidence supports. What I can say, as a reasoned reading rather than an established fact, is that “the model just diagnoses” is an easier story to fund than “the model makes what the pathologist already decides auditable,” and that asymmetry likely shaped where the capital went first.

Nor do I think perception and standardization are mutually exclusive bets. A well-governed foundation model, (inside a validated system, with regulatory oversight, with an explicit measurement purpose, as with AIM-NASH) does produce real clinical value. The question isn’t foundation model yes or no. It’s whether the model was designed to make a human judgment auditable, or to replace it with nothing auditing anything.

What If the Money Chased the Right Question

CAP, RCPath, and the ICCR have spent years working on synoptic reporting protocols and harmonizing histopathology terminology across the U.S., Europe, Australasia, and Asia. They receive a fraction of the capital any new foundation model attracts. And yet, by every regulatory and clinical signal available, they’re the ones actually solving the problem that costs the patient real time.

Foundation models don’t have to sit outside that effort, they can become one of its instruments, rather than its alternative. There’s a concrete agenda here, and pieces of it are already scattered across what these models already do well:

Quantify interobserver disagreement instead of just predicting a label, turn variability itself into a visible metric, not noise to discard. Automatically translate between different classification systems (Japan/West, AJCC/UICC, different Gleason versions) instead of demanding the world converge on one. Audit diagnostic decisions already made, flagging where human judgment departs from a reference consensus, without overriding it. Version diagnostic criteria the way code gets versioned, so you can trace when and why a rule changed. And cross-validate across centers, using the model itself as a thermometer for how much a diagnosis varies by who reads it and where. None of these five is “the model just diagnoses.” All of them use the same technology.

The question isn’t whether foundation models will keep improving. They will. The question that actually matters (the one almost no one is asking before the next funding round) isn’t which model has the best AUROC this week. It’s which clinical decision, currently resting on one person’s judgment in one hospital, becomes more reproducible and more auditable because of what’s being built.

References

Schlemper RJ, Itabashi M, Kato Y, et al. Differences in diagnostic criteria for gastric carcinoma between Japanese and western pathologists. Lancet. 1997;349(9067):1725-1729.
Schlemper RJ, Riddell RH, Kato Y, et al. The Vienna classification of gastrointestinal epithelial neoplasia. Gut. 2000;47(2):251-255.
Chen RJ, Ding T, Lu MY, et al. Foundation models in computational pathology: methods, applications and clinical implications. BMJ Oncology. 2024.
Wang D, Khosla A, Gargeya R, Irshad H, Beck AH. Deep learning for identifying metastatic breast cancer. arXiv:1606.05718 [Internet]. 2016 Jun 18.
PathAI. About PathAI [Internet]. Boston: PathAI; [cited 2026 Jul]. Available from: https://www.pathai.com/about-us
General Atlantic. PathAI Secures $60M in Series B Funding Led by General Atlantic and Existing Investor General Catalyst [press release]. 2019 Apr.
PathAI. PathAI Announces EMA Qualification for AIM-MASH AI Assist (qualified by FDA/EMA under the name AIM-NASH as a Drug Development Tool for MASH clinical trials) [Internet]. PathAI; 2025.
Massachusetts Institute of Technology. 6.S897 Machine Learning for Healthcare, Lecture 12 Notes [Internet]. MIT OpenCourseWare; Spring 2019.
Lawler R. Musk targeting coast-to-coast test drive of fully self-driving Tesla by late 2017. TechCrunch. 2016 Oct 19.
Tesla. Autonomy Investor Day [corporate presentation]. Palo Alto: Tesla; 2019 Apr 22.
Creative Destruction Lab. Machine Learning and the Market for Intelligence [conference]. Toronto: University of Toronto; 2016 Nov.
Bareja R, Carrillo-Perez F, Zheng Y, Pizurica M, Nandi TN, Tian L, et al. A benchmark study of vision and pathology foundation models for computational pathology. Nat Commun. 2026.

The Relocation of Fragility : Innovation doesn't eliminate fragility. It redistributes it.

Dr. Luis Cano — Fri, 24 Jul 2026 10:12:24 GMT

THE MONTHLY REVIEW

I remember the meeting where it started. End-of-month review. Big projects, the company I worked for at the time. Someone announced the multiplex experiment (twelve markers, twelve colors) would be ready in a month. So I prepared. Weeks of it. Protocols, new analysis frameworks, papers on twelve-channel imaging. I wanted to walk in ready the day the data showed up.

It didn’t. One channel, too much noise. Another, nothing. Crosstalk between markers that were never supposed to touch each other. Someone would fix a problem and another one would surface, different channel, different sample, same headache. My analysis protocol sat in a folder. Waiting. For three months.

The experiment that should have taken a month took four.

With a single marker and a single color, the workflow is almost boringly simple: the pathologist looks at the sample, and if the stain came out a little pale, reads it anyway. Notes it, moves on, adjusts the protocol next round. The system tolerates error and keeps working. Twelve colors don’t work that way. Each new channel doesn’t add to the last one, it’s one more point where the whole experiment can break.

For months I chalked it up to bad luck, or a project that was just unusually hard. It wasn’t either. It was something plainer, and honestly more uncomfortable to sit with: every innovation promises to remove a problem. Almost none actually does. What it does, nearly every time, is move the problem somewhere else.

WHAT WE ALREADY KNEW

What I lived through in that lab wasn’t new. Someone had already described it, decades earlier, looking at a completely different problem.

In 1983, Lisanne Bainbridge wrote a short essay on industrial automation that today reads like prophecy. She wasn’t writing about hospitals or labs. She was writing about control rooms, operators in front of factory panels. But she described, forty years ahead of time, the exact mechanism now playing out in precision medicine.

Her argument breaks into two parts. Both matter, equally. First: automate an operator’s routine tasks, and you don’t make that operator unnecessary. You leave them, precisely, the hardest tasks, the ones nobody figured out how to automate. The ones that demand the most judgment, right when the operator has the least practice at exercising it.

The second part is the one that gets under my skin. Pull an operator out of daily control, and you also pull away the constant practice they needed to intervene well the day something breaks. We ask them to passively watch a system that almost never fails, and then react with expert precision at exactly the moment, rare and unexpected, that it does.

Take a pathologist who now reads “positive” or “negative” off a screen. Slide after slide, they’re no longer building the instinct that used to let them clock a strange case on sight. So when the algorithm fails for a reason that has nothing to do with biology (a bad staining batch, say, or a different scanner) that pathologist simply has less recent practice to catch it. Bainbridge never needed to see a single AI model to see this coming.

THE SYSTEM, NOT THE OPERATOR

A year later, Charles Perrow published Normal Accidents, after investigating the Three Mile Island nuclear accident. Bainbridge explains what happens to the human operator. Perrow explains what happens to the system.

Two ideas do most of the work here. Interactive complexity: so many components, tangled up with each other in ways nobody designed on purpose, that one tiny failure combines unpredictably with another tiny failure somewhere else entirely. And tight coupling: rigid sequences, zero slack between one step and the next, no time for a human to catch an error before it’s already spreading.

Put both properties in the same system, Perrow argues, and serious accidents stop being some remote possibility you engineer away. They become “normal.” A property of the system. Not a failure of execution.

And there’s something even more uncomfortable: adding safety layers doesn’t always reduce risk. It often increases it. Every new layer brings its own interface, its own failure mode, and nobody fully understands how they interact.

TWO LEVELS OF THE SAME PHENOMENON

Bainbridge and Perrow aren’t describing two problems. They’re describing the same phenomenon from two different altitudes.

The same move that weakens the human who might catch the failure is, almost always, the same move that makes the failure propagate faster than any human could catch it.

That’s not a coincidence. Both come from the same decision: chase consistency without asking who’s going to absorb the variation that keeps existing, always.

THE MECHANISM

Stripped to its skeleton, the process is always the same:

This sequence doesn’t mention artificial intelligence. It doesn’t mention pathology. It works the same way in an airline, a bank, a nuclear plant. AI didn’t invent this rule. It just makes it visible much faster, compressing into months what took other industries decades.

Variability used to have exactly one place to resolve: the pathologist’s judgment. Today it’s spread across six or seven different links, and none of them absorbs error with the ease a trained eye once did.

Adaptation no longer lives in one person. It’s distributed, and, for the most part, invisible.

THE CASE: DIGITAL PATHOLOGY

I come back to my own territory now. I’m not choosing digital pathology because it’s a special case. I’m choosing it because, of all the industries where this mechanism operates, it’s one we can watch most clearly, with the freshest evidence.

The most revealing part isn’t that models fail obviously, in public, in ways anyone would notice. It’s what they’re learning underneath, while everything looks like it’s working.

THE DATA

Two independent studies, with different data and different teams, trained models on millions of histology images from multiple hospitals, then asked whether those models could guess which hospital an image came from just by looking at its internal representation. In both, the answer was yes, with 88% to 98% accuracy, higher than the same model’s accuracy at identifying the tumor.

The system wasn’t learning biology. It was learning provenance. It’s Bainbridge again: the model became an expert at the easy task, and stayed blind to the hard one nobody taught it well, because nobody knew how.

Once that kind of algorithm becomes the core of a biomarker selecting patients for a Phase III trial, it stops being a technical curiosity.

THE DATA

The Digital PATH project evaluated ten independent algorithms scoring HER2 in breast cancer. On clearly positive cases, consensus was nearly perfect: 97%. In the low-expression range (today’s segment of greatest therapeutic interest) disagreement rose substantially, for the usual reasons: tissue folds, staining artifacts, chromatic variation a pathologist would wave off without a second thought.

GRACEFUL DEGRADATION

A pathologist degrades gradually. Worth pausing on that, because I don’t want it to sound like idealization: pathologists are people, not machines. Nobody delivers excellent performance every single day, across an entire career. What sets their kind of failure apart isn’t that they don’t fail. It’s that they don’t fail all at once.

Faced with an imperfect sample, a pathologist doesn’t just stop working. Doubt shows up first. Confidence drops, a notch at a time. They start asking for a second opinion more than they used to. And almost always, they know it’s happening, they know when they’re seeing worse than they did a few years back, know when the mistakes are creeping up, and more often than not, can tell you themselves when their sharpest years have already passed.

A model doesn’t degrade that way. It holds near-perfect performance right up to the exact instant a sample falls outside what it learned. There, it doesn’t dip, it collapses, with the same numeric confidence it had when it was right. No prior doubt. No threshold sliding slowly. Nobody inside noticing something’s starting to go wrong.

A pathologist knows when they’re reaching their limit, even when that limit arrives with age, not with a hard sample. A model almost never knows it reached its own.

TWO INDUSTRIES, ONE PATTERN

In 2018 and 2019, Boeing wanted to compete faster with the new Airbus A320neo. It fitted larger, more efficient engines onto the classic 737 airframe. The new engine size changed the plane’s aerodynamics just enough that, in certain maneuvers, the nose tended to pitch up on its own.

Instead of redesigning the airframe (expensive, slow, full recertification required) Boeing solved it with a layer of software: MCAS, which pushed the stabilizer down automatically. The system relied on a single angle-of-attack sensor, with no redundancy. When that sensor failed, the software pushed the nose toward the ground, again and again, overriding the pilots. They didn’t know MCAS existed.

Interactive complexity and tight coupling, exactly as Perrow described them: one invisible component that reduced the slack between a faulty sensor and a tragedy to zero.

The second case is more mundane: software infrastructure. A single server fails simply, it goes down, you bring it back up. The industry moved to distributed systems on Kubernetes precisely to eliminate that fragility.

But at scale, documented industry practice shows the same pattern: the fragility didn’t disappear. It relocated into the configuration across dozens of nodes. Small, undocumented variations (configuration drift) produce failures that are hard to isolate, because no single component is broken. The problem lives in the interaction, not the part.

THE QUESTION THAT REMAINS

Nobody in that end-of-month meeting was doing anything wrong. Every decision made sense at the time. That’s the whole point: fragility doesn’t show up because somebody screws up. It shows up because the system gets so consistent, so tightly coupled, that there’s no room left for anyone to be a little bit wrong and just keep going.

AI is going to keep transforming precision medicine. Faster every year. The real challenge of the next decade isn’t building smarter algorithms. It’s building organizations that can absorb the new fragility those algorithms will inevitably introduce, somewhere else in the system, when we’re not looking.

Maybe the real measure of technological maturity isn’t how much we automate, but where we choose to place our next fragility.

References:

Bainbridge L. Ironies of Automation. Automatica. 1983;19(6):775-779.
Perrow C. Normal Accidents: Living with High-Risk Technologies. Princeton, NJ: Princeton University Press; 1999.
de Jong ED, Marcus E, Teuwen J. Current Pathology Foundation Models are unrobust to Medical Center Differences. arXiv preprint. 2025;arXiv:2501.18055v2. See also: Kömen J, Marienwald H, Dippel J, Hense J. Do Histopathological Foundation Models Eliminate Batch Effects? A Comparative Study. Presented at: AIM-FM Workshop, NeurIPS; 2024, independent finding, same pattern.
Salgado R, Allen J, Gallas B, et al. Agreement Across AI-Based Digital and Computational Pathology Platforms Scoring HER2 in Breast Cancer: Results from the Friends of Cancer Research Digital PATH Project. Mod Pathol. 2025;38(3):100944.
Human Factors and Ergonomics Society. Human Factors & Aviation Safety: Testimony to the United States House of Representatives Hearing on Boeing 737-Max8 Crashes. Washington, DC: HFES; 2019.
Industry practice documentation on configuration drift in Kubernetes at scale (platform engineering sources, not peer-reviewed literature).

THE COLD WAR OF DRUG DEVELOPMENT

Dr. Luis Cano — Fri, 17 Jul 2026 09:58:04 GMT

They Never Met

A researcher submitted a manuscript at 2:13 in the morning, lit only by the blue glow of a screen in an empty lab. Six thousand miles away, a few hours later, a committee unanimously approved an acquisition worth several hundred million dollars. Neither one ever knew the other existed.

They worked in time zones that rarely overlapped while awake: one signed emails from a university lab, a hospital’s name printed beneath his signature; the other signed from a high floor overlooking a river that wasn’t his. No shared boss, no shared budget, no shared corporate language.

And yet they ended up fighting for exactly the same thing.

One stockpiled publications, the other stockpiled acquisitions, neither fired directly at the other, and both advanced toward the same target, even if neither could yet name it.

Neither one knew he was at war.

Neither did I.

I understood it much later, once I stopped watching from the outside and found myself walking between both fronts.

This is the story of how I got there (unplanned, unrequested) and why, after crossing the border in both directions, I could never fully belong to either world again.

The Terrain

It took me longer than I’d like to admit to understand what, exactly, both sides were fighting to control. I didn’t see it then. I saw it later, once I’d lived long enough on both sides to recognize the pattern.

For now, this much: it wasn’t the technology, and it wasn’t the talent. It was something quieter, something that had been waiting for both of them for years. And like any contested territory, no one who lingers too long near the border stays neutral for long.

Neither did I. By the time I noticed, I was already walking back and forth across it.

The Side That Taught Me Not to Trust

They sent me to the academic territory first. Nobody taught me the first lesson on purpose. I learned it in a conference room with too many chairs and too little air, watching a colleague, brilliant, honest, exhausted after months of bad sleep, present a model trained in a single hospital, on a single scanner, to a committee that gave him a standing ovation before the last slide finished loading.

Nobody asked what would happen if that same model met a different scanner, in a different hospital, with a different staining protocol. It would have been a reasonable question, and an uncomfortable one. It also would have delayed publication by six months, and in that territory, six months is the distance between being first and being a footnote in someone else’s paper.

At first I thought it was an exception. Then I started listening to the hallway conversations. Over coffee, the question was almost never “does this work?” It was “who else is working on this?” Or “when do we submit?”

That’s when I understood I wasn’t watching different researchers. I was watching an incentive structure doing exactly what it had been built to do.

Like any rookie agent, I’d arrived with the wrong manual. I thought I was entering a world governed purely by scientific curiosity. Instead I found an economy as competitive as any stock market, just denominated in a different currency: the paper, the citation, the h-index, the next grant.

The system wasn’t optimized to produce tools that worked on an ordinary Tuesday, in a hospital with a broken scanner and half the staff out sick. It was optimized to produce a manuscript that could survive peer review before another lab, on another continent, published the same idea a few weeks earlier.

It wasn’t cynicism. It was the incentive structure doing precisely what it was designed to do.

The suspicion became evidence. A letter published in The Lancet, led by researchers at Columbia’s Data Science Institute, audited nearly 2.5 million articles indexed in PubMed and found that AI-fabricated references had multiplied twelvefold in just two years. Nature separately estimated that tens of thousands of 2025 publications may include invalid AI-generated references that were never properly verified.

Nobody designed that outcome on purpose. It’s what happens when an entire system learns to measure production speed before it measures the soundness of what’s produced.

But the speed wasn’t what unsettled me most. It was what the speed was hiding.

Most pathology AI models were born, raised, and proved brilliant inside a single hospital: the same scanner, the same stain, the same workflow. The moment they left, performance often collapsed. The technical literature has an elegant name for it: domain shift. In practice it means something much simpler, the model hadn’t learned to recognize cancer. It had learned, mostly, to recognize the fingerprints of the lab where it was born.

I thought I’d found the enemy.

For a while I believed it was a defect unique to that world: the vanity of publication, urgency dressed up as discovery, a system that confuses speed with truth.

I was wrong. When I crossed to the other side of the wall, I found the exact same defect, wearing a different suit.

The Other Side Said the Same Thing About Us

Nobody in the industrial territory talked about publications. They talked about deployment speed, competitive windows, “we can’t fall behind.” The conference rooms had better coffee and worse views, but the tension in the air was identical to the academic committee’s: the same urgency, in different clothes.

In six months I watched a wave of consolidation with no recent precedent in the history of digital pathology. Roche bought PathAI. AstraZeneca acquired Modella AI (4). Sanofi expanded its alliance with Owkin. A new press release seemed to land every few weeks: another acquisition, another license, another strategic deal.

From the outside, it read like a series of carefully calculated bets. From the inside, it felt different.

I watched entire committees approve investments worth hundreds of millions of dollars without anyone in the room able to answer, first, a surprisingly simple question:

Which specific clinical problem will this purchase solve?

The answer almost always arrived later.

A McKinsey survey put numbers to that feeling: 100% of life sciences leaders had experimented with generative AI. Only 32% had taken concrete steps to scale it. Barely 5% claimed to have gained a real competitive edge from it (6).

A hundred percent arm themselves. Five percent know why.

That’s when I realized I’d been wrong. It wasn’t greed. It was fear, dressed up as strategy, the same fear I’d just left behind, only speaking a different language.

I started keeping a mental list of the phrases I heard on both sides of the border: “We can’t fall behind.” “If we don’t do it, someone else will.” “We’ll sort out the details later.” The words changed. The fear never did.

The Basement

That’s where I understood what the territory actually was. It didn’t happen in a meeting, or in front of a slide deck. It happened in the basement of a university hospital.

Servers hummed under fluorescent lights. On one side, cardboard boxes of biopsies from another era; on the other, the hard drives meant to replace them. Nobody there talked about artificial intelligence. Nobody talked about war.

They talked about space, about budget, about maintenance, about what to do with fifteen years of biopsies nobody had looked at since the day they were archived.

I remember thinking it was an archive. Then I understood I was looking at something else.

Petabytes of human tissue photographed at microscopic resolution, piling up year after year with no one quite sure what to do with them yet, not because they were useful today, but because everyone sensed they’d be valuable eventually.

And then it clicked: that wasn’t an archive. It was the territory both sides had spent years fighting over without ever admitting it.

A digital pathology archive isn’t a medical record. It’s a territory.

And like any territory, the map never belongs to whoever lives on it. It belongs to whoever manages to control it.

What unsettled me most was that neither side knew the patient whose tissue had been sitting there for years, waiting.

That’s when the war stopped feeling abstract. They weren’t fighting over a technology, or even a market.

They were fighting over millions of biological stories turned into data, belonging to people who will likely never know that their biopsy became the ground on which two different worlds learned to compete.

The First Stray Bullet

Every war, even a cold one, eventually claims a first casualty. This one didn’t bleed. It lost credibility.

On April 18, 2026, the New England Journal of Medicine published the case of an 87-year-old man with bronchial casts after inhaling wildfire smoke. The clinical image spread quickly among physicians worldwide. For eleven days, nobody suspected a problem, until an anonymous reader on PubPeer noticed something almost invisible: the numbers on the measuring tape in the image followed a sequence that was mathematically impossible.

Eleven days later, NEJM retracted the article. The authors admitted they’d used an AI tool to reposition the tape measure, to make the image “neater and more legible”.

It wasn’t fraud in the classic sense. It was something more unsettling: well-intentioned authors using a new tool without pausing to ask where aesthetic editing ended and scientific integrity began.

The AI didn’t lie. It simply did exactly what it was asked to do. And maybe that’s the most uncomfortable part of the whole story.

Stray bullets are never fired by one person. They’re fired by an entire system that has learned to move first and ask questions later, trusting that someone else (the reviewer, the committee, the regulator) will correct course before the error travels too far.

This time it took eleven days. Next time it might not.

Because the real casualty was never that image. It was something much harder to rebuild: trust.

And in any cold war, once trust starts to erode, both sides end up losing more than they think they’re defending.

The Weapon Nobody Fires

Not every collaboration in this war is a trap. But even the ones that work run on the logic of deterrence, not trust.

During the Cold War, nobody built a nuclear arsenal hoping to use it. Its real value lay elsewhere: in the other side knowing it existed. The weapon changed the negotiation long before anyone considered firing it.

Something strikingly similar happens in biomedicine.

The alliance between AstraZeneca, Daiichi Sankyo, and Leica Biosystems developed Quantitative Continuous Scoring (QCS), a computational pathology algorithm for scoring HER2. In the DESTINY-Gastric01 trial, it proved able to predict response to trastuzumab deruxtecan, and a few months later the collaboration expanded to a new biomarker, TROP2, in lung cancer.

It’s a real clinical success. Real patients benefited from that work.

But seen through the logic of this war, the algorithm serves another function too. Every new validated biomarker, every new platform, every expansion of the agreement is also a signal: a public demonstration of capability.

It isn’t built only to be used. It’s built so everyone knows it can be used, and like any arsenal, its mere existence changes how everyone else negotiates.

The large public-private consortia play a different role. BIGPICTURE in Europe (10), or the Cancer AI Alliance in the United States, function like the cities where both blocs agree to sit down and negotiate without ever fully trusting each other: neutral ground where hospitals, universities, and pharmaceutical companies share rules before they share advantages.

That isn’t naivety. It’s a recognition of reality: both sides understand that if nobody builds shared infrastructure, neither one gets very far.

Trust remains limited. Cooperation, meanwhile, becomes unavoidable.

And maybe that’s the strangest lesson of any cold war: non-proliferation treaties never meant the powers stopped arming themselves. They meant both sides had realized that some shared rules were the only way to avoid destroying the very thing they were competing for.

The Desertion

There comes a point when a correspondent stops asking who’s winning, and starts asking whether either side can actually win. That moment came for me long after I’d left both behind.

For years I thought I’d crossed from one side to the other. Today I believe I never left the same conflict.

Academia can’t build the next generation of precision medicine without industry’s infrastructure and capital. Industry can’t build it without the scientific credibility, the samples, and the knowledge that only academia produces.

I thought this resembled Mutually Assured Destruction. It wasn’t that. It was something stranger: neither side could destroy the other, because each depended entirely on what the other produced.

The rivalry was real. So was the interdependence. And maybe that’s the most uncomfortable paradox in this whole story.

The enemy never had a name, a logo, or a bank account. It was an incentive structure capable of making intelligent, honest, well-meaning people defend positions that, seen from the border, looked far more alike than either side was willing to admit.

In academia, I learned to distrust industry. In industry, I discovered they distrusted academia exactly the same way. That’s when I understood neither had ever been describing a different enemy: they were just using different languages to name the same problem.

Desertion didn’t feel like a decision. It felt like a loss.

I discovered I no longer fully belonged to either world. I’d learned both languages, but I could no longer speak either one without keeping the other’s accent.

I didn’t choose a side. I lost both, and stayed on the border, the territory almost no one claims, because almost no one yet knows what to do with it.

The Cold War of drug development won’t end when one side wins. It will end when we stop building new walls between them and finally start building the bridges that make the border unnecessary.

For years I thought my job was to cross that border and report what I saw from both sides. Today I believe it’s to help others who no longer need to cross it to understand each other.

Maybe that’s every correspondent’s fate: to stop telling the story of the wall, and start drawing the blueprints for the bridge.

I think that’s the work that comes next.

References:

Topaz M, et al. Fabricated citations: an audit across 2.5 million biomedical papers. The Lancet. 2026 May 9;407(10541):1779-1781.
Nature News. Tens of thousands of 2025 publications may include invalid AI-generated references [Internet]. 2026.
Roche. Roche enters into a definitive merger agreement to acquire PathAI to transform AI-driven diagnostics [Internet]. Basel: Roche Media Release; 2026 May 7.
Modella AI. Modella AI announces acquisition by AstraZeneca to advance AI-driven oncology R&D at global scale [Internet]. Boston: Business Wire; 2026 Jan 13.
Owkin. Owkin to build AI agents as part of a multi-year K Pro collaboration with Sanofi [Internet]. New York/Paris: Owkin Newsfeed; 2026 Jun 5.
McKinsey & Company. Scaling gen AI in the life sciences industry [Internet]. 2025 Jan 10.
Wang Y, Mu X. Retraction: Bronchial casts from inhalation of forest-fire smoke. N Engl J Med. 2026;394:1634. DOI: 10.1056/NEJMc2605962.
Retraction Watch. NEJM retracts case study for AI-manipulated imagery [Internet]. 2026 May 1.
Leica Biosystems. Leica Biosystems announces expansion of collaboration to scale precision medicine and develop AI-powered diagnostics with AstraZeneca and Daiichi Sankyo [Internet]. 2026.
IHI Innovative Health Initiative. BIGPICTURE project factsheet [Internet]. European Union; 2026.
Cancer AI Alliance. Federated consortium for oncology AI research [Internet]. 2025.

Healthcare Didn't Adapt to AI. AI Adapted to Healthcare.

Dr. Luis Cano — Fri, 10 Jul 2026 09:58:08 GMT

I’ve noticed something lately that I hadn’t questioned enough: AI is everywhere now. On the phone, in the search bar, even on the pathology viewer, I still can’t figure out why it's in the refrigerator. We adopt it without asking whether we need it. We know it’s there. We know it’s already part of nearly everything we use. But we rarely stop to ask the uncomfortable question: do we actually know how to use it?

I don’t mean generating a song, an image, or a video in three seconds. That part is easy. I mean something harder: do we know how to recognize when the model is hallucinating? Do we know where it fails, and why? Writing a prompt and getting a convincing answer is not the same as knowing whether that answer is correct.

They sold us AI first with the promise of replacing us, complete with an expiration date on several professions (radiology included). That narrative did plenty of damage on its own. Once it became clear the story was more complicated than advertised, they rebranded it: AI would make us more productive instead… the hype we’re living through now. (you know fragile egos, apparently, find it easier to rebrand than to admit they got it wrong). But whichever narrative gets used, we land at the same bottleneck: AI illiteracy dressed up as technological arrogance, the idea that no learning is required, that typing a prompt is enough.

And if this happens with one person and a chatbot, what happens when the one who never stops to ask is an entire hospital, and the chatbot is an algorithm involved in decisions about patients?

Picture a hospital that buys an AI-based clinical decision support system. The pitch is clean: less administrative burden, faster diagnoses, less friction. Eighteen months later, that hospital has a new AI governance committee, a new algorithmic-bias auditor role, a quarterly recalibration protocol, and a regulatory validation team working overtime. The model works. The hospital, as a whole, is not simpler. It’s more complex than it was before the purchase.

The thesis of this essay is easy to state and hard to accept:

Healthcare’s structure didn’t adapt to AI. AI adapted to it.

THE ORGANIZATIONAL MIRROR

In 1967, computer scientist Melvin Conway submitted a paper to Harvard Business Review. HBR rejected it, on the grounds that his thesis lacked sufficient empirical evidence (great scientific minds at work, evidently). Conway published it the following year in Datamation, and Fred Brooks would later christen it “Conway’s Law” in his classic The Mythical Man-Month.

The law is simple: any organization that designs a system produces a design that is, literally, a copy of its own communication structure. A hospital where radiology, oncology, and pathology run as administrative silos won’t produce a unified clinical AI, it will produce three islands of AI, one per specialty, inheriting exactly the disconnection the algorithm was supposed to solve. And let’s be honest: radiology is already there, and pathology is the discipline right behind it. We have no interconnected algorithms, no shared viewers across specialties, we’re still splitting the screen into four panels every time we want to see a patient’s full picture at once.

But Conway’s isn’t the only theory that explains this. According to Sociotechnical Systems Theory, no technology is successfully adopted if the technical subsystem is optimized without redesigning the social subsystem in parallel, including roles, hierarchies, and workflows. Path Dependency explains why organizations stay anchored to legacy infrastructure even when better alternatives exist, the same mechanism that has kept the QWERTY keyboard alive nearly a century after the Dvorak layout proved more efficient. And DiMaggio and Powell’s Institutional Isomorphism explains why health systems end up building nearly identical AI governance committees: not because it’s the most efficient choice, but because legitimacy (resembling one’s peers) outweighs technical efficiency.

There’s one more piece of that same theory worth naming: decoupling, described by sociologists John Meyer and Brian Rowan in 1977. Their finding is an uncomfortable one, institutions convene committees, draft policies, and publish AI governance frameworks to satisfy what the sector expects of them, while clinical staff, day to day, ignore or quietly adapt those algorithms to keep up with patient demand. The committee exists. The algorithm gets used however the floor actually needs it used.

Each of these theories, from a different angle, says the same thing: technology doesn’t transform the organization. The organization transforms the technology.

A PATTERN WE'VE LIVED BEFORE

This pattern didn't start with artificial intelligence. Economic historian Paul David documented one of the most cited cases in industrial economics: the electric dynamo was already technologically mature by 1890, yet aggregate US industrial productivity stayed flat for the following three decades.

THE DATA POINT

Steam-era factories were built several stories tall, organized around a single central engine that distributed power through a system of shafts and leather belts. When electricity arrived, engineers simply swapped the steam engine for an electric motor, and left the entire building architecture untouched. Productivity didn’t take off until the 1920s, when a new generation of engineers redesigned the factory from scratch: one floor, horizontal layout, one small motor per machine.

Medical informatics repeated the same mistake, twice. Electronic health records weren’t designed around the physician’s cognitive workflow or the patient’s needs, they digitally emulated the structure of a billing-oriented paper chart, and today they’re one of the documented drivers of clinician burnout. PACS, in its first decade of adoption in radiology, reproduced the analog workflow of the lightbox: a digital station simulating film hung on a wall.

The question this essay asks is whether artificial intelligence in healthcare is repeating the same simple substitution for a third time.

THE THREE BUREAUCRACIES

Before the taxonomy, it’s worth seeing what this simple substitution looks like in practice. An AI model built to predict a hospitalized patient’s deterioration (or the onset of sepsis) can perform beautifully on paper. But if the hospital’s alert workflow wasn’t redesigned alongside the algorithm, that model simply floods nursing staff with notifications nobody has the time or the mandate to act on. The result isn’t a safer hospital. It’s one more alert that staff eventually learn to tune out.

Rather than simplifying operations, forcing AI into legacy frameworks generates a new taxonomy of administrative load, deployed across three overlapping layers.

The first two layers, are the most visible and the most discussed in health-innovation circles. The third is the one almost nobody measures, and it’s the most expensive.

Call it cognitive bureaucracy, or more precisely: the displacement of clinical work, from executing diagnostic tasks to constantly supervising and validating an algorithm’s recommendations. Three recent studies, across three different specialties, document the same phenomenon using three different methodologies.

DRATSCH ET AL., 2023,
27 radiologists across three experience levels, reading 50 mammograms with BI-RADS suggestions from a simulated system. When the AI suggested the wrong category, inexperienced radiologists’ accuracy dropped from 79.7% to 19.8% (a 59.9-point fall); the most experienced dropped from 82.3% to 45.5%
BUDZYŃ ET AL., 2025,
Multicenter observational study in Poland, the first real-world (not experimental) clinical evidence of AI-driven deskilling. The adenoma detection rate in standard, non-AI-assisted colonoscopies fell from 28.4% to 22.4% after the same endoscopists were routinely exposed to AI-assisted colonoscopy
ROSBACH ET AL., 2026,
28 pathologists estimating tumor cell percentage, with and without AI assistance. In 7% of cases, a pathologist reached a correct independent diagnosis and then changed it to accept an incorrect algorithmic estimate

This is the cognitive debt no committee audits, because it never shows up in any regulatory file. It lives in the time every clinician spends (without anyone formally asking for it) deciding whether or not to trust what the algorithm just suggested.

THE STRONGEST COUNTERARGUMENT

It would be dishonest to make this case without facing its best counterexample. Ambient AI scribes (systems that transcribe and draft the clinical note in real time during the visit) are, so far, the strongest evidence that AI can genuinely cut administrative load.

THE COUNTEREXAMPLE

A multicenter quality study evaluated the rollout of an ambient AI platform across 263 physicians and advanced-practice providers in six US health systems. After 30 days of regular use, self-reported burnout dropped from 51.9% to 38.8%, and after-hours documentation time fell by an average of 0.90 hours per day

The result is real. But conceptually, it doesn’t invalidate the thesis, it displaces it. The time AI frees up from writing notes is rarely protected institutionally for rest or for a more careful visit. Quite the opposite: more free time is supposed to translate into more revenue, and under financial pressure, hospital administrations tend to reabsorb that time immediately, adjusting schedules to demand a higher volume of visits per hour (13). The physician, meanwhile, stops being the author of the clinical note and becomes the critical editor of an AI-generated draft, a new role, with a different kind of legal responsibility, that introduces its own form of fatigue.

Even the proposals designed specifically to fix automation bias run into the same problem. An emerging model called AI-Instigated Human Oversight (AIHO) flips the logic: instead of a human passively supervising the algorithm, the model itself assesses its own uncertainty and decides when to escalate a case to a human expert. It’s conceptually elegant. But implementing it in practice requires the hospital to stand up a new committee, one dedicated to monitoring the performance of the AI validator that supervises the original AI. One more governance layer, stacked on top of the ones that already existed.

The pattern holds even in the case that best contradicts it, and even in the case explicitly designed to fix it: the technology frees up time or reduces bias, and the organization, without redesigning itself, finds a way to fill that space right back up with a new structure.

THE DIGITAL PATHOLOGY CASE

No field illustrates this dynamic more clearly than digital pathology itself.

The technical burden here has a very concrete size: a mid-sized lab processing 200,000 slides a year needs roughly 300 additional terabytes of storage annually just to keep the high-resolution image archive alive. That’s not a cost anyone brings up when pitching the business case for the scanner.

The Royal College of Pathologists in the UK documented a full department-wide validation of a histopathology department under its whole slide imaging guidance. Overall diagnostic concordance was excellent, just 1.3% net discordance between digital image and glass slide (49 of 3,777 cases). But the time it took individual pathologists to complete their own validation process varied wildly within the same institution: one pathologist focused on prostate biopsies finished in 10 days; another, in gynecologic pathology, took 793.

That variability has nothing to do with the technology. It’s about how each organization treats regulatory compliance, as a box-checking exercise in legal conformity, or as a real opportunity to redesign how pathologists actually work.

Memorial Sloan Kettering Cancer Center offers the counterexample of what happens when redesign actually occurs. Its digital pathology rollout, documented in Archives of Pathology & Laboratory Medicine, cut physical slide requests from the department archive and an off-site surgical center by 93% and 97% respectively, and confirmatory immunohistochemistry orders by 30%, with projected savings above $267,000 a year from staffing and archive-service restructuring.

The difference between 10 days and 793 isn’t in the scanner. It’s in whether the organization decided, from the outset, to redesign itself around the technology or simply bolted it onto what it already had.

THE ACTION FRAMEWORK

Back to the uncomfortable question we opened with. It isn’t only whether we know how to use AI individually, whether we can recognize when it hallucinates, when to trust it and when not to. It’s the same question, scaled up: does an organization know how to use AI, or does it only know how to buy it?

None of the theories covered earlier, Conway’s Law, Path Dependency, Institutional Isomorphism, decoupling, suggests the problem goes away with a better model. They suggest the opposite: that the problem persists until the organization stops asking “which AI use case should we implement now?” and starts asking “which part of our structure needs to change for this to actually work?”

It’s the same distinction, ultimately, as the one between writing a prompt and knowing what to do with the answer. If this organization were being founded today, with AI already available from day one, would we design it the same way?

That’s the question worth asking before bed, obviously, not while you’re doomscrolling.

References:

Conway ME. How do committees invent? Datamation. 1968 Apr:28-31.
Carayon P, Hundt AS, Karsh BT, Gurses AP, Alvarado CJ, Smith M, et al. Work system design for patient safety: the SEIPS model. Qual Saf Health Care. 2006;15(Suppl 1):i50-8.
David PA. Understanding the economics of QWERTY: the necessity of history. In: Parker WN, editor. Economic history and the modern economist. Oxford: Blackwell; 1986.
DiMaggio PJ, Powell WW. The iron cage revisited: institutional isomorphism and collective rationality in organizational fields. Am Sociol Rev. 1983;48(2):147-60.
Meyer JW, Rowan B. Institutionalized organizations: formal structure as myth and ceremony. Am J Sociol. 1977;83(2):340-63.
David PA. The dynamo and the computer: an historical perspective on the modern productivity paradox. Am Econ Rev. 1990;80(2):355-61.
Downing NL, Bates DW, Longhurst CA. Physician burnout in the electronic health record era: are we ignoring the real cause? Ann Intern Med. 2018;169(1):50-1.
Reiner BI. The insidious problem of fatigue in medical imaging practice. J Digit Imaging. 2013;26(5):788-92.
Dratsch T, Chen X, Rezazade Mehrizi M, Kloeckner R, Mähringer-Kunz A, Püsken M, et al. Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology. 2023;307(4):e222176.
Budzyń K, Romańczyk M, Kitala D, Kołodziej P, Bugajski M, Adami HO, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10:896-903.
Rosbach E, Ammeling J, Ganz J, Bertram CA, Conrad T, Riener A, Aubreville M. Stuck on suggestions: automation bias, the anchoring effect, and the factors that shape them in computational pathology. Mach Learn Biomed Imaging. 2026;3:126-47.
Olson JE, et al. Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Netw Open. 2025;8(10):e2534976. doi:10.1001/jamanetworkopen.2025.34976.
Nundy S, Cooper LA, Mate KS. The quintuple aim for health care improvement: a new imperative to advance health equity. JAMA. 2022;327(6):521-2.
Optum Advisory. Ambient AI medical scribes: efficiency gains, burnout uncertainty, and governance risks. 2026.
Sirintrapun SJ, et al. A lifecycle framework for digital pathology implementation: integrating infrastructure, workflow, regulatory compliance, and AI readiness for sustainable adoption. J Clin Transl Pathol. 2026.
Royal College of Pathologists. Department-wide validation in digital pathology: experience from an academic teaching hospital using the UK Royal College of Pathologists’ guidance. Diagnostics. 2023;13(13):2144.
Hanna MG, Reuter VE, Samboy J, England C, Corsale L, Fine SW, et al. Implementation of digital pathology offers clinical and operational increase in efficiency and cost savings. Arch Pathol Lab Med. 2019;143(12):1545-55.

The Sophistication Trap : Why More Technology Doesn’t Necessarily Create More Value

Dr. Luis Cano — Fri, 03 Jul 2026 09:59:08 GMT

The hospital had done the transition right. On paper.

They bought the high-resolution scanner. Set up cloud storage for gigapixel images. Rolled out digital viewers on every pathologist’s desktop. In the board presentation, the project looked exactly as it should: digital pathology, end to end, ready to scale.

Six months later, the scanner had become the most expensive paperweight the hospital ever bought. It sat idle most of the week. Of every thousand images that were supposed to reach the cloud, barely four hundred did. Technicians weren’t complaining about the technology itself, they were complaining that it added a step instead of removing one. Day to day, in the hallways, people kept doing what they’d always done: looking through the microscope.

One day, a resident had a hard case and wanted a second opinion. He found the attending in the hallway: “The image is ready in PACS, everyone can log in whenever you’re free.” “Do you have the slides?” the attending asked. “Yes, right here, I just finished scanning them,” the resident said. “Let’s go to the five-head scope. Call the others.” Thirty-year-old technology was still the first instinct and it worked better, faster, with less friction than any digital workflow the hospital had built that year.

That scene has repeated, with variations, in more than one hospital I’ve seen up close. And every time, the question left hanging in the air isn’t “why aren’t they using what we bought?” It’s an earlier question, one nobody asked before buying: was the problem we had actually the one this technology solves?

The question almost everyone asks when evaluating health tech is the wrong one. It’s not which is best. It’s what works, what fits, what actually solves the problem.

No hospital buys thinking about the sharpest image resolution or the most advanced algorithm on the market. Or it shouldn’t. It buys what fits the workflow that already exists. What an exhausted technician can operate at six in the evening without extra friction. What solves the real problem — not the problem the vendor described in the demo, under the best light, in the best-case scenario, in front of the committee that approves the budget.

This is a story about that gap. About how the health industry and the pharma industry around it, has learned to confuse two questions that sound alike but aren’t: how sophisticated is this technology, and what problem does it actually solve in production? The evidence, not just the anecdote, suggests that gap is wider, and far more expensive, than most organizations are willing to admit.

For months I thought the problem was artificial intelligence. It wasn’t. It’s a much older pattern that AI just makes more visible, and more expensive. Over time I started plotting the projects I saw (the scanner gathering dust, the five-head scope beating fiber optics) on two very simple axes. I didn’t know yet that it would turn into a matrix with a name of its own. We’ll get to it, with the data behind it.

Subscribe now

THE NUMBERS DON’T LIE

The wall of evidence

For a while I treated the five-head-scope scene as a nice anecdote, the kind of story you tell in a conference hallway and forget by the next day. But I went looking for whether it was an outlier, and after going through dozens of reports and studies on AI adoption in health care, I landed on an uncomfortable conclusion: it wasn’t. It’s the local version of a pattern that industry-wide evidence confirms with a consistency that surprises even people who expected it.

RAND CORPORATION, 2024

80% of AI projects fail, double the failure rate of IT projects that don’t involve AI.

RAND’s study isn’t a market estimate or an attention-seeking headline. It’s the result of structured interviews with sixty-five data scientists and engineers, each with at least five years building AI models in industry or academia. Its most uncomfortable finding isn’t technical: the most common cause of failure isn’t a broken algorithm. It’s a problem that was never properly defined in the first place.

The gap between investment and real value shows up in every serious report published on the subject in the last two years. Boston Consulting Group’s October 2024 report, “Where’s the Value in AI?,” surveyed a thousand C-suite executives across fifty-nine countries. Its headline finding: despite $252.3 billion in aggregate spend in 2024, 74% of companies couldn’t demonstrate tangible value from their AI investments.

In health care, the pattern sharpens. Qventus’s 2026 report, built on interviews with more than sixty technology, clinical informatics, and AI leaders at U.S. health systems, documents a tension any clinical leader will recognize instantly: 65% rated board pressure to operationalize AI a 7 or higher on a 10-point scale. 94% said delaying adoption would put them at a direct competitive disadvantage. And yet only 4% of those same health systems have actually scaled an AI solution with measurable clinical or operational results.

Board pressure: 65%. Feeling of competitive disadvantage from waiting: 94%. Systems that actually scaled with measurable results: 4%.

The gap isn’t a problem of insufficient ambition. It’s almost the opposite: there’s plenty of ambition, and what’s missing is rigorous problem selection before budget and reputation get committed to the solution.

The case closest to the five-head-scope scene, though, comes from a study published in July 2025 in the Journal of the American Medical Informatics Association, which surveyed forty-three leading U.S. health systems across thirty-seven distinct AI use cases. The finding is almost a mirror of Act I: imaging and radiology are, by far, the clinical area with the highest AI deployment, 90% of organizations report at least partial deployment, but only 19% of those same organizations report a high degree of real success with those tools. In clinical risk stratification, perceived success drops to 38%. The only category with consistent success, 53%, is automated clinical documentation: precisely the application with the least algorithmic sophistication and the least clinical decision authority of the set.

The pattern, taken as a whole, is clear: the closer an AI tool gets to the center of clinical decision-making, and the more sophisticated its architecture, the wider the gap between deployment and perceived success. This isn’t a case of technology that doesn’t exist or doesn’t work in the lab. It’s a case of an organization that never checked, before buying, whether that was actually where the bottleneck lived.

When extreme sophistication loses to extreme simplicity

Health care has been running a natural experiment for more than a decade without calling it that: two bets, nearly opposite in scale, aimed at the same kind of problem (reducing preventable harm to patients) with results that should be required reading in any innovation committee.

The sophisticated bet

IBM Watson for Oncology started from a seductive premise: if an AI system had beaten the best human players on Jeopardy!, why couldn’t it process the world’s oncology literature and generate treatment recommendations more rigorously than any individual oncologist? IBM acquired Truven Health Analytics for $2.6 billion and Merge Healthcare for $1 billion to consolidate the training data it needed. Aggregate internal investment topped $4 billion. The collaborative project with MD Anderson Cancer Center alone cost more than $62 million before it was cancelled.

The system never operated with the reliability it promised. Independent evaluations documented clinically unsafe recommendations, including suggesting bevacizumab for lung cancer patients with severe hemoptysis, a contraindication any first-year oncologist would catch immediately. In concordance studies against tumor boards in South Korea, the system reached just 77% agreement in breast cancer and 54.5% in gastric cancer, far below what’s needed to sustain clinical trust.

The root cause wasn’t a lack of computational power. It was a model trained almost exclusively on Memorial Sloan Kettering’s prescribing patterns, marketed as a universal validation system, with no transparency into how it reached its recommendations, the classic black-box problem, and no real adaptation to the local clinical and epidemiological contexts where it was deployed.

The simple bet

In parallel, a team led by surgeon and researcher Atul Gawande designed an intervention that fit on a single sheet of paper: a nineteen-item surgical safety checklist, organized around three moments (before anesthesia induction, before incision, before the patient left the operating room) read aloud by the surgical team itself.

MULTINATIONAL STUDY, EIGHT HOSPITALS

−36% in major surgical complications after introducing the checklist (from 11.0% to 7.0%).

In-hospital deaths fell more than 40%, from 1.5% to 0.8%. Infrastructure or licensing cost was, in practice, zero. The barrier wasn’t technical, it was cultural: surgeons who initially saw the checklist as a bureaucratic step slowing them down.

The same pattern shows up at a smaller scale inside pathology itself. An improvement program at a histology lab, published in the American Journal of Clinical Pathology, combined a barcode system with a Lean redesign of the microtomy workflow (no new equipment purchased) and achieved a 78.64% reduction in labeling errors, along with a 35.28% drop in reprocessing time.

$4 billion and a black box recommending unsafe treatments, against a sheet of paper read aloud that saved measurable lives. The difference wasn’t the budget. It was whether anyone asked, before starting, where the problem actually lived.

Neither case exists to argue that technological sophistication is never worth it. Digital pathology has its own legitimate high-value quadrant, and we’ll name it explicitly further down. The point is more precise, and more uncomfortable: sophistication and value are independent variables. They can move together. But just as often they split apart, and when they do, almost no organization catches it before the contract is signed.

THE VILLAIN

It’s not the technology. It’s the expectation we built around it.

It would be comfortable and wrong, to close this argument by saying health care simply buys bad technology. That’s not what the evidence shows. Paige Prostate Detect’s algorithms work. Mitosis-detection models work. Three-dimensional radiotherapy dose planning works, and it’s exactly the kind of problem, high mathematical dimensionality, impossible to solve by hand, where sophistication is a necessity, not vanity.

The villain of this story isn’t the algorithm. It’s the expectation built around it: the idea, almost never said out loud but present in every innovation committee, that not having a visible AI strategy is itself a strategic risk, regardless of whether the problem the organization actually faces requires that level of sophistication.

And it’s worth saying plainly: this pattern didn’t start with AI, and it won’t end with it either. Health care and pharma have lived it before, under other names. The ERP wave of the nineties and 2000s promised to integrate the entire hospital operation and, more often than not, ended in underused systems kept alive out of fear of reversing the investment. The Big Data craze promised that more data, on its own, would produce better clinical decisions. Blockchain was going to solve health record interoperability and, in the vast majority of hospital pilots, never made it past the innovation committee. AI is simply the newest, most expensive version of a question health care has spent thirty years avoiding: is the problem we have organizational, or technological? Because only one of those two gets solved by buying something.

We jumped into the pool because everyone around us was jumping. Did anyone bother to check if the pool had water in it?

That expectation has a name in the organizational behavior literature, and it’s worth naming precisely, because a problem without a name is much harder to manage in a board meeting.

Complexity bias

The human brain tends to assume that intricate systems (the ones that demand more cognitive effort to understand) are inherently superior to direct approaches. Research on clinical behavior around language models shows that both residents and specialists are vulnerable to this bias: their diagnostic accuracy drops when they discard a simple, correct answer in favor of an elaborate, sophisticated one generated by an intelligent system, just because it sounds more rigorous.

Innovation theater

Adopting cutting-edge AI carries enormous external symbolic value. It lets a board project technological leadership to investors, regulators, and high-profile patients. When the primary goal becomes announcing the acquisition of the system, not auditing, months later, the real value it generated for the patient or the technician operating it, innovation stops being a bet and becomes a performance.

Escalation of commitment

Once an institution commits a large initial investment (financial and reputational) to an AI platform, whoever made that call tends to justify further investment down the same path, to avoid the psychological cost and public scrutiny of admitting a strategic loss. The project doesn’t stop when it stops making sense. It stops when there’s no more budget (or career) left to protect.

The incentives of whoever decides

A CIO’s or CMIO’s professional standing is often tied to the size and technical sophistication of the systems portfolio they run. Approving a Lean Six Sigma workflow redesign doesn’t carry the same weight on a corporate résumé as leading the rollout of “autonomous AI agents”, even when the former has a better shot at moving the number that actually matters.

The system was designed, without anyone planning it that way, to reward announcing the AI project, not auditing, six months later, whether that project actually solved anything.

Naming the pattern isn’t new, the literature on tech solutionism, innovation theater, and escalation of commitment has existed for years, in health care and beyond. What’s still rare is an organization pausing, before approving the budget, to ask which of these mechanisms is driving its own decision. Naming the pattern hasn’t prevented it. What’s missing isn’t more literature about the problem. It’s a filter used before signing, not a diagnosis written after.

Sophistication is not an attribute of the solution. It is a property of the problem.

Put differently: every problem has an optimal level of sophistication, not the maximum available, not the easiest to justify in a boardroom, but whatever the problem actually demands. Below that level, the solution falls short. Above it, the solution gets fragile, expensive, and hard to maintain, without buying any additional value. That’s the principle behind the tool that follows.

Share Beyond the Slide

THE FRAMEWORK

The Complexity-Value Matrix

If the villain is the expectation, the defense can’t be an opinion about whether AI is “worth it” in the abstract. It has to be a sequence of concrete questions, applicable before the budget is committed, not a damage audit after the scanner is already gathering dust. This is the Complexity-Value Matrix: the tool I kept sketching, almost without noticing, every time a project started looking more like the scanner from Act I than the five-head scope.

First, here’s what a problem that actually deserves sophistication looks like

Before the sequence of questions, Quadrant I deserves its due, it exists, it’s real, and it isn’t a marginal exception. A pathologist with a light microscope can’t calculate, by eye, the spatial distribution of sixteen hundred morphological biomarkers in a single H&E image. A medical physicist can’t manually optimize beam-incidence angles for radiotherapy to minimize damage to a glioblastoma without compromising surrounding healthy tissue, that’s a high-dimensional geometric problem no heuristic shortcut can handle. In both cases, algorithmic sophistication isn’t a signal for the board. It’s the only way to make visible something the human eye structurally cannot see. That’s the standard any other project should be measured against before approval: does the problem genuinely exceed human cognitive capacity, or does it just feel more comfortable to treat it as if it did?

This is the sequence I use with the teams I work with, four questions that capture what separates Quadrant I from Quadrant II on the map below.

Question 1 : Is the root cause computational or organizational?

Is the bottleneck a pattern the human eye can’t process (mathematical density, data volume, dimensionality) or is it a problem of coordination, communication, or workflow between people? If the root cause is organizational, no amount of algorithmic sophistication will fix it. The surgical checklist and the five-head scope share the exact same diagnosis: the problem was never computational.

Question 2 : If it is computational, does a simple solution already solve most of it?

Before building or buying a complex model, it’s worth benchmarking it against the simplest clinical heuristic available. In predicting 30-day hospital readmissions, published comparisons show the LACE index (four variables, calculated by hand) performs on par, statistically, with deep neural networks, and in some cases beats models like XGBoost, which tend to overfit the noise in local clinical data. If the simple solution already captures 80% of the value, that’s the starting point — not the finish line.

Question 3 : If you go sophisticated, can you name the real value metric before signing?

Not the model’s technical performance metric, sensitivity, specificity, AUC on a curated dataset. The operational or clinical metric you’ll measure in production, against a documented baseline, six months after deployment. If no one in the room can name that metric before approving the project, what’s being purchased isn’t a solution. It’s the feeling of having bought one.

The cost that never shows up on the quote

There’s another reason to take Question 3 seriously: the price on the quote is never the real cost of sophistication. Every additional degree of complexity comes with a bill that shows up nowhere in the contract, staff training, integration with the existing clinical workflow, ongoing model maintenance, governance of the data feeding it, periodic revalidation against performance drift, and the almost never budgeted cost of the cultural change required before anyone in the ER trusts what the machine recommends. None of that shows up in the demo slide. All of it shows up, late, in next year’s budget. A project that looks cheap on the quote and turns expensive in operation didn’t fail on price, it failed because nobody added up the cost of complexity before signing.

Question 4 : What evidence, within six months, would prove us wrong?

This is the question almost no organization asks, and it’s the most important of the four. It’s not about defining success, Question 3 already does that. It’s about defining, in advance, the concrete evidence that would trigger walking away. Without that threshold written down before starting, the escalation of commitment from Act IV has a clear runway: any ambiguous result gets read as “needs more time,” never as “this isn’t working.”

These four questions can be plotted as two independent axes, technological sophistication and real problem-solving value, producing four quadrants. Placing a project there, before approving it, forces into the open what otherwise stays implicit in the excitement of a demo.

The Complexity-Value Matrix

The most useful thing about this matrix isn’t classifying finished projects, that’s easy, and it changes nothing. It’s using it in the room where the budget gets approved, before the money is committed, and forcing someone to say out loud which quadrant they think the project will land in, and why. The evidence gathered here suggests a good share of today’s digital health investment ecosystem sits, without anyone deciding it explicitly, in Quadrant II: high sophistication, low value.

Quadrant II projects rarely fail because the algorithm is bad. They fail because they try to solve an organizational problem with a computational tool.

HOW WE BEAT IT

The job the technology actually has to do

There’s a conceptual framework, originally developed by Clayton Christensen for consumer markets, that translates surprisingly well to an operating room or a pathology lab: Jobs to Be Done. Its premise is that people and organizations don’t buy products for their technical attributes. They “hire” them to do a specific job, in a specific circumstance.

The real job of a pathology chief isn’t operating the most advanced algorithm on the market. It’s cutting diagnostic turnaround time and reducing interobserver variability on hard cases. The real job of a hospital CIO isn’t having an “AI strategy” to present to the board. It’s making sure resources (people, compute, budget) produce better, measurable, sustained clinical decisions per patient.

When the job gets defined independently of the solution, something the “buy the most advanced system” logic systematically hides comes into view: in plenty of cases, a change in physical workflow design (where the five-head scope sits, how the report sign-off rotation is organized, who sees the urgent case first) achieves the same result, at a fraction of the transaction cost, without the adoption curve that quietly kills so many well-intentioned AI projects.

The organization that wins this decade won’t be the one that announces the most ambitious AI project. It’ll be the one that knows, before approving the budget, which quadrant it’s going to land in.

This isn’t an argument against sophistication. It’s an argument for earning it. Quadrant I exists, it’s real, and the organizations that occupy it (in computational pathology, precision radiotherapy, biomarker discovery) didn’t get there by buying the most advanced thing available. They got there after ruling out, with evidence, that a simpler solution could solve the same problem. Sophistication, when it’s earned, stops being a signal for the board and becomes what it was always supposed to be: the right answer to a question someone actually bothered to ask first.

THE FUTURE THAT ACTUALLY WORKS

The question that changes the conversation in the room

Next time a vendor walks into a committee room with a flawless demo of an autonomous agent system, or a foundation model promising to solve the hospital’s diagnostic bottleneck once and for all, there’s one question worth more than any sensitivity or specificity figure on the slide: if the problem is organizational, why are we evaluating a computational solution?

You don’t need to reject the technology to ask that question. You just need to refuse to jump into the pool before checking there’s water in it. Every organization has its own five-head scope waiting in some corner, its own thirty-year-old solution that still outperforms last year’s sophisticated project. The question was never whether it exists. It was whether anyone bothered to look for it before buying something new.

Sophistication is not an attribute of the solution. It is a property of the problem.

Before approving any AI project

This is the short version of everything above, built to print and bring to the room.

Before approving your next AI project, it's worth spending an hour answering these four questions. It's usually a lot cheaper than finding out, six months later, that the project landed in the wrong quadrant. If you want to run the Complexity-Value Matrix against your next investment, or one you've already made that isn't delivering, I'd be glad to help, just send me an email to luiscanoayestas@gmail.com.

References:

Ryseff J, De Bruhl BF, Newberry SJ. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI. Santa Monica (CA): RAND Corporation; 2024. Report No.: RR-A2680-1.
Boston Consulting Group. Where’s the Value in AI? Boston: BCG; 2024 Oct.
Qventus. Beyond the Pilot: How CIOs Are Operationalizing AI Across Health Systems in 2026. San Francisco: Qventus; 2026 Apr.
Poon EG, Lemak CH, Rojas JC, Guptill J, Classen D. Adoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges. J Am Med Inform Assoc. 2025 Jul;32(7):1093-1100.
Flach RN, van Dooijeweert C, van Diest PJ, et al. Prospective Clinical Implementation of Paige Prostate Detect Artificial Intelligence Assistance in the Detection of Prostate Cancer in Prostate Biopsies: CONFIDENT-P Trial. JCO Clin Cancer Inform. 2025.
Strickland E. IBM Watson, heal thyself: how IBM overpromised and underdelivered on AI health care. IEEE Spectr. 2019;56(4):24-31.
Haynes AB, Weiser TG, Berry WR, et al. A surgical safety checklist to reduce morbidity and mortality in a global population. N Engl J Med. 2009;360(5):491-499.
Heher YK, Chen Y, Pyatibrat S, Yoon E, Goldsmith JD, Sands KE. Achieving high reliability in histology: an improvement series to reduce errors. Am J Clin Pathol. 2016;146(5):554-560.
Christensen CM, Hall T, Dillon K, Duncan DS. Competing Against Luck: The Story of Innovation and Customer Choice. New York: HarperBusiness; 2016.
Kim H, Ock M, Park JY, et al. Real-world concordance rate of Watson for Oncology with multidisciplinary tumor board decisions in South Korea. BMC Cancer. 2023.

Hacking the Medical System: How Necessity Drove Change But Not Adoption

Dr. Luis Cano — Fri, 26 Jun 2026 09:54:27 GMT

I remember the first iPhone I bought. It was 2014. The novelty wasn’t just the hardware, it was what lived inside it: applications, connectivity, the feeling of holding a computer in your hand. I started learning Swift, Apple’s programming language, convinced that everything could eventually fit inside that device.

I also remember something else from those years: the jailbreak. Apple had its official developers, but there was a whole community of enthusiasts who knew how to break the code and build their own software. Ringtones. Games. System customization. Small technical transgressions, full of enthusiasm and irreverence. The official ecosystem said this far. The user said we’ll see.

That moment matters because it captures something fundamental about closed systems: when a platform solves the wrong problem, users build parallel solutions. Always. Not out of malice. Out of necessity.

Now imagine that dream reaching medicine. Apps arriving in the clinic. Heart rate. Respiratory pattern. Fundoscopy. Dermoscopy. Histological analysis. An App Store for physicians where every diagnostic problem has a downloadable solution.

The question isn’t whether there would be a jailbreak. The question is how many. And what each one reveals about the problem the industry chose not to see.

Medicine adopted the model, and it made sense

The “Appleification” of diagnostic medicine wasn’t an accident or an arbitrary imposition. It was a logical response to a real problem.

When AI began showing promising results in medical image analysis (detecting pulmonary nodules, classifying biopsies, quantifying biomarkers) the industry hit an immediate technical wall: variability. An algorithm trained on Hamamatsu scanner images doesn’t necessarily perform the same on Aperio images. Hematoxylin-eosin staining varies between labs. Image compression, scanning parameters, color calibration: all of it introduces noise that degrades model performance.

The only practical way to control that variability, in 2016 or 2018, was to control the hardware. It was the right answer to the problem of the moment.

That’s how closed ecosystems were born. The scanner manufacturer also provided algorithms validated on its own images. Buy the machine, get the models. Give us your data, we control the variability. The hospital gets reproducible results. The manufacturer gets lock-in. It was (and in certain contexts still is) the only technically viable architecture.

The medical App Store wasn’t born from greed. It was born from engineering. And that is precisely the hardest kind of error to correct: the one backed by good reasons.

What if the problem was never detecting more tumors?

There is an assumption the medical AI industry adopted so early and so naturally that it never had to be stated as one. It became invisible. And invisible assumptions are the most dangerous kind.

The assumption was this: the bottleneck in diagnostic medicine is the diagnosis.

If physicians miss tumors, catch them too late, or classify them with too much inter-observer variability, then the solution is an algorithm that detects, classifies, and quantifies more accurately. The logic is sound. The premise is the problem.

THE PARADOX : The more accurate diagnostic AI became, the less clear it was that diagnosis was medicine’s main problem.

A radiologist in a high-pressure hospital environment reads approximately 34 CT scans per day. Each CT generates an average of 679 reconstructed images. With 5.5 net hours of reading time available, that leaves roughly 0.86 seconds per image. In that context, adding one more window the radiologist has to open, confirm, and reconcile with the PACS isn’t an improvement. It’s additional friction on a system already running at its limit.

And while the industry was building detection algorithms, radiologists were spending up to 20% of their working day preparing for and attending multidisciplinary tumor boards. Pathologists were using 2.4 hours of preparation time for every hour of scheduled meeting. Nobody had an algorithm for that. Nobody was building one.

The real bottleneck

20% of radiologists’ and pathologists’ working day at high-complexity centers goes to preparing and attending multidisciplinary meetings, with zero AI support.

The question the industry never asked out loud is the simplest one: what does the physician do with their time when they’re not interpreting images? And could AI help with that?

The answer was already written. In the jailbreaks clinicians had built for themselves.

150 algorithms, 60 developers and silence

September 2025. Bayer announces it is discontinuing Calantic Digital Solutions and winding down operations at Blackford Analysis, its radiology AI subsidiary.

It wasn’t a surprise to those who knew how to read the signals. It was, like the García Márquez novel, a death everyone saw coming and nobody could (or chose to) prevent.

The scale of the collapse

~10% of global radiology AI platform spending eliminated overnight.

The numbers are concrete. Blackford had built one of the most ambitious ecosystems in the sector: over 150 third-party applications, over 60 developers, a promise of being the neutral App Store that no hardware manufacturer could be. Bayer had acquired it in 2023 (barely two years earlier) convinced the future ran through there.

The official statement spoke of “strategic reprioritization” and “reinvestment in growth areas.” The real translation is more direct: the model wasn’t working. Hospitals were buying. Hospitals were piloting. Hospitals weren’t using.

The immediate fallout was dramatic. The 60 developers who had built their products on the platform suddenly had no distribution channel. Hospitals that had begun integrating those workflows had to evaluate emergency replacement paths. CIOs who were already notoriously cautious about long-term AI commitments now had one more reason to be.

And the market kept moving. Because markets always keep moving. Signify Research documented that in the same month Bayer exited, three additional acquisitions were announced in the sector adding to the seven already registered that year. Consolidation didn’t wait for the body to be buried.

The sector’s confession

41% of radiologists feel that AI tools developed within their organization don’t adequately address their real-world needs.

Four out of ten radiologists. Involved in the development of those apps. Saying the result doesn’t serve what they actually need. The assumption the industry never questioned (that the bottleneck was diagnosis) had been written in that number since May. Nobody wanted to read it.

The dead: 60 developers without a platform. The wounded: hospitals locked into ecosystem contracts they can’t exit because switching costs are structurally prohibitive. The survivors: those who understood early enough that the value wasn’t in the algorithm catalog. It was in removing friction.

Three jailbreaks. One law…

When a system solves the wrong problem, users build parallel solutions, always. And the solutions they build are the most honest evidence of the problem the system ignored.

What remained after the marketplace collapse wasn’t a vacuum. It was a parallel ecosystem that filled the spaces the official model left uncovered. Three jailbreaks. Three manifestations of the same phenomenon. Three diagnoses of the same error.

JAILBREAK I (the simple fix) : WhatsApp, ChatGPT, and the workaround everyone uses

The first jailbreak required no specialized technical knowledge. Everyone did it. Everyone still does.

Digital imaging systems were extraordinarily good at one thing: seeing. High-resolution viewers. Instant zoom. Stain synchronization. From the standpoint of pure diagnostic visualization, digital pathology and radiology PACS represent decades of sophisticated engineering. Sharing was a different story.

Sharing a biopsy image with a colleague at another hospital. Sending a complex case to an expert in another country at 10pm. Discussing an ambiguous finding with the oncologist before Monday’s tumor board. Those communication flows weren’t in the original system design.

So physicians did what people always do when the system doesn’t solve their problem: they broke it with whatever they had at hand. WhatsApp became the de facto teleconsultation channel in dozens of countries. ChatGPT started appearing in radiology report drafting workflows. Excel continued managing worklists where the PACS couldn’t reach.

Here is the insight that changes everything: none of those workarounds is a diagnostic hack. WhatsApp doesn’t detect tumors. ChatGPT doesn’t classify biopsies. Excel doesn’t perform cell segmentation. They are all productivity hacks. Communication hacks. Coordination hacks.

Physicians didn’t jailbreak the system to get better diagnoses. They jailbroke it to save time. That is exactly the opposite of what the industry had assumed they needed.

JAILBREAK II (the sophisticated fix) : Open-source platforms, Ki-67, and the medical Cydia

The second jailbreak required more, technical knowledge, time, and some audacity. Few people did it. But what they built has consequences that extend well beyond those who use it.

While major manufacturers were constructing validated, certified, regulated diagnostic algorithm ecosystems, a parallel market was taking shape at the margins of the official system. In open-source histological image analysis platforms (developed in academic environments and distributed at no cost) biomarker quantification algorithms began circulating. Ki-67, estrogen and progesterone receptors, cell segmentation tools published as scientific papers with available code, ready to deploy.

Who was using them? Private clinics without enterprise system budgets, Anatomical pathology labs with limited resources. Physicians with their own practices who needed to quantify a biomarker to make a treatment decision and couldn’t wait months for the official system to provide that capability. Each one under their own responsibility. Without formal regulatory validation, without institutional traceability, but working.

This is medical Cydia. The alternative repository that emerged when the official App Store didn’t offer what the user needed.

This jailbreak is also evidence of the same phenomenon, but with a nuance that cannot be romanticized.

An incorrect Ki-67 quantification can influence a chemotherapy decision. An algorithm without formal validation can produce systematically biased results across certain patient populations or staining protocols. The risk is not hypothetical. It is structural.

Initiatives have also emerged proposing that specialists themselves (without data scientists on the team) develop and distribute image analysis models. The intention is noble. The execution misses something fundamental: the difference between an algorithm that works on the training dataset and one that works in real clinical practice requires exactly the expertise being proposed for elimination. The Cydia jailbreak on iPhone could leave your phone unusable. The medical equivalent is worse. Here the operating system that can break is not the iPhone.

JAILBREAK III (the functional apocalypse): The invisible infrastructure that can be wrong

The Proximity Paradox of Digital Pathology

Dr. Luis Cano — Fri, 19 Jun 2026 09:58:54 GMT

A few weeks ago, during a consulting session, I was talking with a client about emerging trends in hospital operations when the conversation took an unexpected turn. He told me the next big problem in hospitals was going to be a space problem.

Space? I asked. Hospitals are already large. He smiled. Not space to grow, he corrected me. Space to reorganize.

Of course. Now there’s more free space in the pathology area. Before, you needed one or two pathologists per office. But now with digital pathology, everyone is diagnosing from home. The offices are empty. We need to figure out what to do with all that room.

He said it with the ease of someone describing a storage reorganization. No alarm. As if it were the natural consequence of a sound technology investment.

And technically, it is. Digital pathology works. Scanners are precise, viewers are fast, pathologists can work from anywhere. The hospital bought technology to solve a problem, and the technology solved it. The issue is that no one had asked whether that was the only problem that needed solving.

We built software to eliminate distance. What we didn’t see is that in doing so, we created a different kind.

The Invisible Infrastructure of Diagnosis

I remember being a pathology resident in Peru. Telepathology didn’t exist yet (if you wanted to be precise, we had barely discovered the word) but what did exist were hallways and the five-headed microscope.

Every time I had a doubt about a case (and I had many) I picked up my slide, stood up from my microscope, and walked down to the office of the senior pathologist responsible for that area. I knocked. I said: “Do you have a couple of minutes? Can you look at this?”

He would look. Sometimes he asked: what’s your diagnosis? I answered. He confirmed, or didn’t, and thirty seconds later I was back at my microscope with something I hadn’t had before: certainty. Other times, what started as a simple question turned into twenty or thirty minutes at the double-headed microscope, or the five-headed one, if other residents were rotating through, or if the case was too good not to share.

That walk was not in any protocol. It did not appear in any procedure manual. Nobody had designed it.

But it was, simultaneously:

Real-time quality control
Implicit knowledge transfer
Diagnostic calibration across generations
Mentorship without a name
Departmental culture building

All at once. And nobody was measuring it.

When Microsoft researchers analyzed the behavior of more than 61,000 knowledge workers during the shift to remote work, they found that collaboration networks became more static and fragmented. Communication did not disappear. It simply stopped crossing organizational boundaries.

Messages increased.
Spontaneous conversations declined.

This is not a pathology problem. It is what happens to any knowledge industry when informal proximity disappears.

Today, that conversation requires opening a formal consultation ticket in the system. Or sending a message on Teams. Or waiting for the next tumor board.

What used to take three minutes and happen spontaneously now requires planning, credentials, and aligned calendars.

Subscribe now

The Right Solution for the Wrong Problem

Digital pathology was built with a clear objective: eliminate geographic distance. And on that objective, it delivered with precision.

I think about the time I received a set of tissue blocks from Puno nearly six weeks late. Andean geography is not a metaphor: it is a biopsy that takes weeks to reach a specialist, or never does. In that context, digital pathology is the exact answer to the exact problem it was designed to solve.

The error was not in the technology, the error was in applying that same solution (generic and transactional), to contexts where the problem was radically different. A pathologist in a European hospital does not need the geographic distance of the Andes eliminated. They need the informal proximity to the colleague three doors down preserved.

That is where the generic solution became the problem.

The evidence is beginning to show that what looked like generational nostalgia may be something more substantive. When researchers studied how departments changed after digitization, they found a consistent pattern: residents, technicians, and support staff described a progressive fragmentation of the interactions that had structured their daily learning.

What disappeared was not an office. It was an invisible network of micro-consultations, corrections, and spontaneous conversations.

The paradox: we solved geographic distance and created collegial, institutional, and mentorship distance

The first generation of platforms was built on a transactional logic:

Radiology Has Been Here Before

Pathology is not the first diagnostic field to live through this transition. Radiology did it two decades earlier. And the scars are documented.

At Brown University, radiologists tracked how their daily face-to-face interactions with clinical colleagues fell from roughly 500 to 600 per day in the early 1980s to near zero by the mid-1990s. The department shifted from a vibrant social hub to a dark, isolated reading room.

The consequences were not only cultural. They were economic, legal, and clinical. Burnout rates among radiologists reached between 54% and 72%. Malpractice claims involving teleradiology show patient mortality rates of 36%, compared to 20% in traditional in-person radiology. The commoditization of reading, professional isolation, and disconnection from the hospital transformed a medical specialty into a transactional, outsourced service.

Radiology didn’t have time to design its transition. Pathology still does.

Not Everything Can Be Digital

The argument is not to return to glass slides. The argument is more precise: there are activities where physical co-presence creates value that no screen has managed to replicate without degrading clinical or educational quality.

Digital pathology does not have to choose between efficiency and culture. It has to design when each applies.

The Incomplete Diagnosis

We need to do a complete macroscopy of the problem. We had a distance, we had technology, we added them up and found an answer. That was the dominant framework. But it cannot remain so, because the moment has come to add a layer where we solve not only the logistics but also the human dimension, which is the fundamental pillar of medicine.

Industry played a decisive role: they heard about our problem and solved it. But we were never in the same room looking at the full picture together. If we had been, we would have realized the problem was incompletely framed from the start.

Hospitals bought technology to eliminate geographic distance. They did not buy technology to preserve collegiality, because nobody told them that also needed to be bought. And industry did not include it in the solution because nobody asked.

That is the incomplete diagnosis. And it is precisely where hospitals and industry need to sit down together, before signing the next contract.

There are questions that do not appear in any digital pathology procurement process today:

How does this platform support urgent informal consultations without requiring a formal consultation ticket in the LIS?
Is there a real-time synchronous viewing environment that allows a faculty member and a resident to interact on the same image from different locations?
How does the system preserve departmental culture when its professionals stop sharing physical space?
What contingency architecture ensures diagnostic continuity during a network outage?

These questions do not have a single answer. They have a contextual one. And that is precisely the conversation that is not yet happening systematically between those who buy the technology and those who design it.

There are emerging models pointing in the right direction: synchronous co-navigation platforms, hybrid affinity spaces, task-oriented presence policies. But they are isolated initiatives, not systems.

Building that system requires a real partnership. Not a sale. Not an implementation. A co-construction that begins by asking the right question: what problem are we actually solving?

The future of digital pathology is not a choice between the physical and the digital. It is the deliberate design of systems that preserve the benefits of both.

Empty offices are not just a space problem. They are the symptom of an incomplete diagnosis. And getting the diagnosis right is, precisely, what we are supposed to do.

A closing question:

If you were designing a pathology department from scratch today, what would you decide to keep in person?

References:

Hanna, M. G. & Ardon, O. Digital pathology systems enabling quality patient care. Genes Chromosomes Cancer 62, 685–697 (2023).
Debating digital pathology? Consider the risk of not doing it. https://resources.ironmountain.com/whitepapers/d/debating-digital-pathology-consider-the-risk-of-not-doing-it.
News and press releases. Sectra Group https://sectra.com/news-and-press-releases/.
Bringing digital pathology to the hospital environment. https://healthcare-in-europe.com/en/news/bringing-digital-pathology-to-the-hospital-environment.html.
Pearson, S. D., Moreno, R. & Trnka, Y. Informal Consultations Provided to General Internists by the Gastroenterology Department of an HMO. J Gen Intern Med 13, 435–438 (1998).
Krupinski, E. A. Virtual slide telepathology workstation of the future: lessons learned from teleradiology. Hum Pathol 40, 10.1016/j.humpath.2009.04.011 (2009).
Ashraf, N. et al. Incidence and factors associated with burnout in radiologists: A systematic review. Eur J Radiol Open 11, 100530 (2023).
Fratt, L. Hot Topic! Has PACS Ignited an Epidemic of Burnout? https://radiologybusiness.com/topics/healthcare-management/medical-practice-management/hot-topic-has-pacs-ignited-epidemic-burnout (2013).
A Digital Pathology Platform Trusted by Hospital Worldwide. https://www.tribun.health/a-digital-pathology-platform-trusted-by-hospital-worldwide.
Browning, L. et al. Impact of the transition to digital pathology in a clinical setting on histopathologists in training: experiences and perceived challenges within a UK training region. J Clin Pathol 76, 712–718 (2023).
Sebastian, M. et al. Applications and challenges of utilizing digital pathology and AI-enabled workflows in clinical trials. J Pathol Inform 20, 100542 (2026).
From Glass to Cloud: The Next Leap for Pathology. The Pathologisthttps://thepathologist.com/issues/2025/articles/september/from-glass-to-cloud-the-next-leap-for-pathology.
Jahn, S. W., Plass, M. & Moinfar, F. Digital Pathology: Advantages, Limitations and Emerging Perspectives. J Clin Med 9, 3697 (2020).
Work & well-being. Microsoft Research https://www.microsoft.com/en-us/research/project/work-well-being/.
Microsoft has found we’re working longer and collaborating less during COVID-19. World Economic Forumhttps://www.weforum.org/stories/2021/10/microsoft-study-covid19-work-hours/ (2021).
Larsen, E. P. et al. Optimizing Radiology Reading Room Design: The Eudaimonia Radiology Machine. J Am Coll Radiol 18, 108–120 (2021).

The Most Expensive Experiment in Modern Biology

Dr. Luis Cano — Fri, 12 Jun 2026 10:00:15 GMT

Picture the moment you receive the results.

Eight months have passed. The budget ran out weeks ago, but the team kept going because the science justified it. The platform performed flawlessly, no incidents, no contamination, no sample loss. Controls passed. The bioinformatics pipeline ran without errors. The heatmaps are spectacular. The clusters appear exactly where they should.

The team is satisfied. Legitimately satisfied.

Then someone in the meeting asks the question nobody expected:

What decision changes if this cluster exists?

Nobody answers.

Not because the answer is complicated. Not because the team is incompetent. But because nobody asked that question when the study began.

The data is real. The biology is interesting. And yet, the study cannot support a single concrete program decision, not a go/no-go, not a patient stratification criterion, not a threshold for the next clinical trial.

Eight months. Hundreds of thousands of euros. What remains is a paper.

> 6 months

the average time committed to a spatial omics program before a team discovers its data cannot support any concrete R&D decision

The villain is not the technology. It is a way of thinking

This scenario is not exceptional. It is the most recurring, and most costly, failure mode in spatial transcriptomics and proteomics programs across pharma and biotech.

And when it happens, the instinct is to look for a technical fault. Review the pipeline. Question the platform. Ask whether the gene panel was right.

But the failure is not there.

It lies in a premise that teams rarely articulate out loud, but that implicitly drives study design:

We have access to a powerful new technology.
Let’s generate data.
We’ll see what we find.

That is technology-first thinking. And it is not a spatial omics problem, it is an organizational pathology that appears in AI, in foundation models, in single-cell sequencing, in high-dimensional proteomics, in any technology impressive enough to substitute for the question.

The technology is often more powerful than the questions teams ask it.

Spatial omics simply makes this especially visible because the investment is large, the timeline is long, and the data is so rich that it is easy to confuse its abundance with its utility.

The tissue architecture nobody saw before the clinical trial

Technology-first thinking does not only produce data without decisions. In oncology, it can produce clinical failures that a well-designed study would have prevented.

Two cases illustrate the problem precisely:

Roche developed a bispecific antibody that required the simultaneous physical co-presence of tumor cells (DR5+) and stromal fibroblasts to trigger apoptosis. Selected patients expressed both targets abundantly on global biopsy. Selection criterion: met.

The clinical trial was halted for lack of efficacy.

Post-hoc spatial analysis revealed that the tissue had a dense tumor core surrounded by fibroblasts strictly segregated to the stromal periphery. Both targets existed. They were never in contact. The tissue architecture made the drug’s mechanism of action impossible, a finding that early spatial transcriptomic profiling would have detected before the clinical investment.

Pfizer designed a bispecific anti-CD3/anti-P-Cadherin antibody to force an immunological synapse between circulating T cells and tumor cells overexpressing P-Cadherin. The target was present. The mechanism was sound on paper.

Resistance did not come from the target. It came from stromal collagen barriers that physically excluded T cells from the tumor compartment, and from niches enriched with suppressive granulocytes that nullified the drug’s lytic activity.

A drug can be biochemically optimized, possess high target affinity, and still fail clinically if the tissue microenvironment's architecture prevents the necessary cellular encounter.

In both cases, the correct question was not “does this tissue express the target?” It was “in which cell, in which compartment, at what distance from the relevant effector cells?” That question distinguishes an actionable study from one that generates impressive heatmaps.

The conversation that should happen before the machine is turned on

Before getting into the framework, it is worth naming something explicitly.

These questions are not really questions about spatial omics. They are questions about how organizations learn. Spatial omics simply makes the problem visible because it is expensive, slow, and produces enormous quantities of information. But the same pattern appears in artificial intelligence, in single-cell sequencing, and in any technology capable of generating more data than an organization knows how to convert into decisions.

The bottleneck is no longer technological capacity. It is the organizational capacity to formulate questions that the technology can answer usefully.

If the problem is pre-experimental, the solution is too. Not a pipeline problem. A design problem, and a governance problem.

The central principle is straightforward: any spatial omics study should begin by defining what concrete decision, about a candidate, a target, patient stratification, a trial design, will change depending on what the data shows.

In practice, this means establishing a Decision Trigger before the experiment begins: an explicit criterion that defines what course of action is taken based on the profiles obtained. Without that trigger, the study does not support decisions, it accumulates exploratory evidence that never integrates into governance processes.

The five questions a spatial omics study must answer before the first sample

This principle can be operationalized into five questions, plus a question zero that is the prerequisite for all the others. These are not a technical post-hoc checklist. They are questions that must be answered unanimously before processing the first sample.

Building an interdepartmental governance committee, with pathologist, translational scientist, bioinformatician, and program coordinator, that does not allow the first sample to be processed until these questions are answered is not bureaucracy. It is the only way to transform a high-dimensional technology into a predictable tool for translational medicine.

Share Beyond the Slide

The bottleneck is no longer the technology

Spatial omics is not overhyped. Its analytical capacity is genuinely extraordinary. What is underdeveloped is the organizational discipline that should precede it.

And this is not a spatial omics problem. It is the pattern that emerges every time a sufficiently powerful technology reaches an R&D team before that team has developed the muscle to formulate the right questions.

We see it in clinical AI. In single-cell. In foundation models applied to pathology. In multimodal proteomics. The constant is always the same:

The bottleneck is no longer technological capability. It is our capacity to formulate questions that technology can answer in a useful way.

Teams that develop that capacity do not only avoid wasting resources. They design studies that yield pharmacodynamic Decision Triggers, biomarkers with regulatory pathways, patient stratification criteria for Phase II trials.

Teams that don’t generate interesting papers.

If this pattern feels familiar, the experiment that generated magnificent data that nobody knew how to bring to a portfolio review meeting, there is a conversation to be had, feel free to contact me.

Suggeested lectures:

Kathryn. How spatial biology improves clinical trial success in oncology. BioLizard https://lizard.bio/how-spatial-biology-improves-clinical-trial-success-in-oncology/ (2026).

Beyond the Slide

Autopsy #002: Two Bodies in the Molecular Window

The Orphan Question

The comfortable hypothesis

The questions are 35 years old

The orphan question

Archaeology, not design

When someone signs

The pool

The questions, returned

Can AI Make Us More Human?

Does it give us time back?

Where does that time go?

What are we optimizing?

What technology never solved

What actually makes care human?

Will we remember?

Further Reading

Autopsy #001 : Oral semaglutide in early Alzheimer's disease

The FDA’s Framework Measures How Autonomous the AI Is, Not Whether the Doctor Is Still Thinking.

The terrain

How the tool shapes the decision

Who can actually control its behavior

The diagnostic library

The questions still open

Further Reading

Who Needs Whom? AI and Pathologists: A Love Story

The dinner and the question

What clinical practice actually provides

A concrete case: who negotiates, who participates

What we should be asking when we’re invited to collaborate

Further Reading

Credential Theater: Should a pathologist learn to code?

The Perfect Precedent

What the Scale Actually Says

Two Places Where It Should Show

The Question That Loses Its Footing

Credential Theater

What We Still Don’t Know

The Questions a Candidate Should Ask

The Right Question

Recommended Reading

The diagnosis no one can sign: Digital Pathology's global blind spot

A decade solving the same problem

The line that changes everything

Not just Peru, and the obstacle isn’t technical

The pattern has a name, and we’ve seen it before

Two caveats before jumping to conclusions

The question the industry still isn’t asking

Further Reading

The Dark Genome and the Two Clocks of Innovation

The First Slide

Why Glioblastoma Wins

The Dark-Genome Hypothesis

What EVX-05 Claims, and What’s Been Shown

Same Day: EVX-05 In, EVX-03 Out

The Precedents That Show Where Prediction Fails

Two Clocks: Biological Validation and Corporate Necessity

What Evidence Would Actually Change Our Confidence

Closing: Back to the Microscope

Confirming Isn't Curing: What a canceled oncology program, an AI CEO, and a student on GPT-4 have in common

How a Person Becomes a Variable

When the Variable Works, and When It Stops Being Enough

We’ve Seen This Movie Before

AI as an Epistemic Amplifier

Who Decides What Counts as Cured

An Autopsy of Four Drugs: The Cause of Death for Four AstraZeneca and Novartis Oncology Programs

What the Family Tells Us

A Simple Diagnosis That Seems Sufficient

The Fourth Case: Lu-NeoB

Where the Chain Breaks

An Experimental Treatment as a Possible Solution

The Warning Sign

When Target Expression Isn't Enough

THE SCENE

55% vs. 0%

THE QUESTION THAT SHOULDN’T MAKE SENSE

TWO OBJECTS THAT SHARE NOTHING

WHY TISSUE MATTERS MORE THAN THE REPORT ADMITS

55% vs. 0%