Bottom Line

Iron deficiency is a disease of iron stores, and anemia is its late stage. By the time hemoglobin and MCV fall, the patient has usually been iron deficient for months, often with symptoms we attribute to stress, sleep, or age. A ferritin below 30 ng/mL makes the diagnosis in a patient without inflammation; in inflammatory conditions, ferritin can sit well above 30 despite empty stores, so transferrin saturation (TSAT) below 20% carries the diagnosis instead. Hepcidin, the liver hormone that controls iron entry into the blood, explains much of the treatment. Each oral iron dose raises hepcidin for about a day, which blunts absorption of the next dose; that is the logic behind alternate-day dosing. Inflammation keeps hepcidin persistently high, which is why oral iron so often fails in chronic kidney disease, heart failure, inflammatory bowel disease, and cancer, and why IV iron, which bypasses the gut, works there. And the diagnosis is never the end of the work: iron deficiency in a man or a postmenopausal woman is gastrointestinal blood loss or malabsorption until proven otherwise.

Why this review

I am preparing for a podcast discussion next week, and this JAMA review by Michael Auerbach, Thomas DeLoughery, and Jennifer Tirnauer kept coming up [1]. It is a narrative review, not a systematic one; the authors say so in their limitations section, and they did not formally grade the evidence. That matters when we read its recommendations. Some rest on randomized trials, some on observational associations, and some on expert consensus. I have tried to label each below.

What makes the review worth a post is its framing. It is about iron deficiency, not iron-deficiency anemia. That distinction changes who we test, what threshold we use, and what we tell the patient whose hemoglobin is normal.

How common is it?

More common than most of us were taught. In NHANES data from 2017 to 2020, about 14% of US adults had absolute iron deficiency, and 11% of adults without anemia, heart failure, chronic kidney disease, or pregnancy still met the definition [1,2]. In a Canadian cohort of more than 600,000 reproductive-age, nonpregnant women, 38% had iron deficiency without anemia and 13% had iron-deficiency anemia [1]. In one Irish cohort followed through pregnancy, iron deficiency rose from 21% at 15 weeks to 84% at 33 weeks [1].

For every patient with iron-deficiency anemia, several more are iron deficient with a normal hemoglobin. Those are the patients a CBC alone will never find.

The mechanism in one paragraph: hepcidin

Iron enters the blood through one door, a transporter called ferroportin, found on intestinal cells that absorb dietary iron and on the macrophages and liver cells that store and recycle it. Hepcidin, made in the liver, binds ferroportin and causes its degradation, closing the door [1]. Hepcidin falls when stores are low, opening the door wider. It rises with adequate stores, with recent iron intake, and with inflammation [1]. (I discussed the inflammation side of this, anemia of inflammation, in an earlier post on hepcidin.) For absolute iron deficiency, two consequences matter: a recent oral dose partly closes the door for the next one, and an inflamed patient may not open it at all.

Ferritin behaves separately. It is an acute-phase reactant, driven up by inflammatory cytokines regardless of how much iron the body holds [1]. That is why ferritin is diagnostic when low and much less informative when “normal” in an inflamed patient.

Symptoms we miss

Fatigue, exercise intolerance, irritability, depression, and difficulty concentrating are the common complaints, and none is specific [1]. Two symptoms deserve to be asked about directly because they are both common and suggestive:

• Pagophagia (ice craving). In 987 consecutive patients with iron deficiency (mean hemoglobin 10.4 g/dL), 40% had pagophagia [3].

• Restless legs syndrome. In the same series, 40% had restless legs, and 22% had both [3]. Restless legs is also the symptom with the best randomized evidence that iron helps: a meta-analysis of 10 trials (455 patients) found that iron reduced the International Restless Legs Syndrome score by 3.55 points compared with placebo, on a 0 to 40 scale [4].

For the broader symptoms such as fatigue and poor concentration, the review presents associations, mostly observational. That is enough to justify testing a symptomatic patient. It is not proof that iron will fix every complaint, and patients should hear that.

Hemoglobin and MCV are late findings. In a pregnancy study, a hemoglobin below 11 g/dL and an MCV below 80 fL each had a sensitivity of only 30% for iron deficiency in the first trimester, and half of 1,749 iron-deficient pregnant individuals had neither anemia nor microcytosis [1,5]. A normal CBC does not exclude iron deficiency.

The diagnosis: why 30, and when 30 is not enough

Ferritin below 30 ng/mL had a sensitivity of 92% and a specificity of 98% for absent bone marrow iron, the historical reference standard, in the study the review cites [1,6]. Low ferritin has essentially one cause: iron deficiency [1].

The problem is the reference range on your lab report. Many laboratories set the lower limit of normal for ferritin well below 30. Martens and DeLoughery have argued that these ranges were derived from populations that included many iron-deficient women, so the “normal” range absorbed the disease it should detect [7]. A ferritin of 18 flagged as normal is not normal; it is iron deficiency. The 30 ng/mL threshold itself is expert consensus, not a value derived from outcomes [1].

Thirty is not universal

The review uses 30, but others draw the line higher.

• 45 ng/mL: The AGA guideline uses a ferritin below 45 to diagnose iron deficiency in patients with anemia, citing a sensitivity of 85% and a specificity of 92% from its systematic review. It calls the right threshold in patients without anemia uncertain [10].

• 50 ng/mL: The argument for 50 comes mostly from physiology. Galetti and colleagues pooled stable-isotope absorption studies in 624 healthy young women and found that iron absorption began to rise once ferritin fell below about 50 [21]. In other words, the body starts conserving iron at that level. A French randomized trial enrolled nonanemic, menstruating women with unexplained fatigue and a ferritin below 50; oral iron reduced fatigue more than placebo, but modestly, by about 3.5 points on a 40-point scale [22].

The argument also runs the other way. Using NHANES data from about 7,500 nonpregnant women, Mei and colleagues found that hemoglobin and soluble transferrin receptor stopped changing at a ferritin of about 25 [23]. By that measure, iron-restricted red cell production begins closer to 25 than to 50.

These numbers answer different questions. Around 25 to 30, stores are running out and red cell production starts to suffer. Around 50, the body has begun to conserve iron. None of the studies supporting 50 used bone marrow iron or clinical outcomes as the reference. Our suggested approach, which is clinical judgment rather than a validated rule: below 30, diagnose iron deficiency and look for the cause. Between 30 and 50, in a symptomatic patient, especially a menstruating woman or someone with restless legs, check TSAT and consider a trial of iron. The review makes a similar point, recommending TSAT whenever iron deficiency remains a concern despite a ferritin above 50 [1].

Inflammation changes the rules. In IBD, CKD, heart failure, and cancer, ferritin rises as an acute-phase reactant, and an iron-deficient patient can have a ferritin of 50 or 80. The review notes that with absolute iron deficiency and coexisting inflammation, ferritin rarely exceeds 100 ng/mL [1,8]. In an inflamed patient, check TSAT; a TSAT below 20% supports iron deficiency even with a ferritin in the “normal” range. The review also recommends checking TSAT whenever iron deficiency remains a concern despite a ferritin above 50 [1].

Two practical points about TSAT. Serum iron rises after iron-containing food or supplements, so TSAT should not be drawn within 5 to 9 hours of ingesting them; fasting is not otherwise required [1]. And a low TSAT with a high ferritin can also mean functional iron deficiency, the iron-restricted erythropoiesis of inflammation, where stores exist but cannot be released. Distinguishing that from combined absolute and functional deficiency is not always possible from the numbers alone.

Finally, the therapeutic trial is a diagnostic test. A reticulocyte response within about a week of starting iron confirms that iron was limiting red cell production; no response suggests another or an additional diagnosis [1].

Find the cause

The review is emphatic, and I agree: diagnosing iron deficiency without asking why is incomplete work [1].

In premenopausal women, heavy menstrual bleeding is the leading cause; the review suggests asking about changing pads or tampons more than every 3 hours or overnight, and passing clots larger than an inch, and evaluating for fibroids, polyps, and bleeding disorders such as von Willebrand disease when bleeding is heavy [1].

In men of any age and in postmenopausal women, the gut is the default. An AGA technical review pooled 18 studies of 12,040 asymptomatic men and postmenopausal women with iron-deficiency anemia: upper endoscopy found an upper GI cancer in 2.0% and colonoscopy found a lower GI cancer in 8.9% [1,9]. The review describes these rates as roughly 100-fold higher than those found in general population screening. The AGA guideline recommends bidirectional endoscopy for these patients even without symptoms [10]. Note that these data come from patients with anemia; the yield in nonanemic iron deficiency is less well defined.

Malabsorption deserves its own checklist:

• Celiac disease: 3.2% of patients with iron-deficiency anemia in a systematic review [1,11]. The review and the AGA recommend testing patients with iron deficiency and upper GI symptoms [1,10].

• Helicobacter pylori and autoimmune gastritis: both impair absorption; the review’s algorithm includes antiparietal cell antibodies alongside celiac serologies and H. pylori testing [1].

• Proton pump inhibitors: gastric acid helps absorb iron, and a meta-analysis of observational studies found a relative risk of iron deficiency of 2.56 with PPI use [1,12]. That is an association; I could not find absolute risks in the review.

• Bariatric surgery: in 58 women studied before and after sleeve gastrectomy or Roux-en-Y bypass, absorption of heme iron fell from 24% to 6%, and nonheme iron from 11% to 5% [1].

Aspirin and NSAIDs, regular blood donation, and intense athletic activity round out the list [1].

Treatment: oral iron, and how hepcidin changed the schedule

Oral iron remains first-line for most patients [1]. Ferrous sulfate 325 mg contains about 60 mg of elemental iron [1]. Avoid enteric-coated and timed-release preparations; they reduce absorption and can stretch treatment to 2 years or more [1].

For decades, guidance favored daily iron divided into two or three doses, on the assumption that more frequent dosing meant more absorption [13]. The hepcidin story upended that. Stoffel and colleagues gave iron-depleted young women 60 mg of iron either daily for 14 days or on alternate days for 28 days, so both groups received the same 14 doses [13]. Hepcidin was higher with daily dosing, and fractional absorption was higher with alternate-day dosing (21.8% vs 16.3%), for total absorbed iron of 175 mg vs 131 mg [13]. Their second trial found that splitting a dose into two doses a day did not improve absorption [13]. These were small, short absorption studies using isotope-labeled iron, not outcome trials.

The clinically relevant test came later. In a masked, placebo-controlled trial of 150 iron-depleted premenopausal women, daily iron for 90 days and alternate-day iron for 180 days produced nearly identical ferritin (43.8 vs 44.8 ng/mL), with significantly fewer GI adverse effects on alternate days [1,14]. Same iron, same result, fewer side effects, but it takes twice as long. For a symptomatic patient, that trade-off is worth discussing.

Absorption helpers and blockers [1]:

• Take with vitamin C or meat protein to improve absorption.

• Avoid tea, coffee, calcium, and fiber within 1 hour before and after the dose.

• Diet alone cannot repair an established deficit, which the review estimates at about 1,000 mg [1].

GI side effects are the main reason patients stop: in a meta-analysis of 20 placebo-controlled trials, adverse effects were more than twice as likely with ferrous sulfate as with placebo [1].

Check that it is working. In iron-deficiency anemia, hemoglobin should rise by at least 1 g/dL within about 2 weeks; a pooled analysis of five trials used that early response to identify patients who would not respond to oral iron [1,15]. No rise despite adherence should prompt a search for malabsorption or ongoing blood loss [1].

When to go straight to IV iron

The review lists the indications [1]:

• Intolerance of oral iron despite schedule or formulation changes

• Impaired absorption: celiac disease, bariatric surgery, autoimmune gastritis

• Chronic inflammatory conditions: CKD, heart failure, IBD, cancer, rheumatologic disease

• Ongoing blood loss greater than oral iron can replace

• The second and third trimesters of pregnancy

• Need for surgery soon

Six formulations are available in the US, and the review states that studies show no difference in efficacy among them. Low molecular weight iron dextran, ferumoxytol, ferric carboxymaltose, and ferric derisomaltose can each replace most of a deficit in one or two infusions of 15 to 60 minutes; iron sucrose and ferric gluconate require multiple visits [1].

The safety story is better than its reputation

Many of us learned to fear IV iron because of high molecular weight iron dextran, which is no longer marketed [1]. True anaphylaxis occurs in roughly 1 in 200,000 infusions [1]. A meta-analysis of 103 randomized trials comparing IV iron with oral iron, placebo, or no iron found no increase in serious adverse events [1,16].

What does happen, in about 1% to 4% of patients, is a minor infusion reaction: flushing, chest pressure, back or chest tightness, without wheezing, stridor, periorbital edema, or hypotension [1]. These are thought to reflect complement activation, sometimes called the Fishbane reaction [1,17]. The review’s advice is to stop the infusion, wait for the reaction to resolve, and restart more slowly; these reactions rarely recur. It specifically advises against treating them with epinephrine and diphenhydramine, which can add tachycardia and somnolence to the picture [1].

The review also reports that diphenhydramine premedication was associated with more minor reactions (38.6% vs 1.7%) [1,18]. I would read that number cautiously. In the source study, an adverse event was defined by the administration of diphenhydramine, epinephrine, famotidine, or hydrocortisone within 24 hours of the infusion, and premedication was presumably given more often to patients judged at higher risk [18]. Both features would inflate the association. The conclusion that premedication is unnecessary is reasonable; the 23-fold difference should not be taught as a measured effect of diphenhydramine. (This critique is mine, based on the study’s methods, not a claim in the review.)

Hypophosphatemia: the formulation difference that matters

Ferric carboxymaltose causes hypophosphatemia, which can be prolonged and produce fatigue, muscle weakness, and cramps [1]. In the PHOSPHARE-IBD randomized trial of 97 patients with IBD, hypophosphatemia occurred in 51% after ferric carboxymaltose and 8.3% after ferric derisomaltose [1,19]. The review recommends checking phosphorus before a second ferric carboxymaltose dose and holding it if phosphorus is low [1]. Readers should know that Dr. Auerbach disclosed fees from Pharmacosmos, which makes ferric derisomaltose [1]; the PHOSPHARE-IBD result is a randomized comparison and stands on its own.

After the infusion

Recheck hemoglobin, ferritin, and TSAT about 4 weeks after IV iron, and give more if ferritin remains below 30 ng/mL or TSAT below 20% [1].

Transfusion is not iron therapy

Red cell transfusion is for severe anemia (hemoglobin below 7 g/dL) with life-threatening bleeding, cardiac ischemia, or hemodynamic instability [1]. It raises hemoglobin immediately, but the review notes that it does not adequately replete iron stores [1]. A transfused patient with iron deficiency still needs iron.

What the review does not settle

• Screening asymptomatic, nonanemic people. Recommendations vary. The European Hematology Association recommends screening a long list of at-risk groups, including menstruating individuals and adults over 65; the American Society of Hematology is still developing a guideline [1]. For pregnancy, the USPSTF found insufficient evidence to recommend for or against screening for iron deficiency in the absence of anemia, while FIGO and the EHA recommend screening all pregnant individuals [1,20].

• Whether treating nonanemic iron deficiency improves outcomes beyond restless legs. The symptom associations are real, but the randomized evidence for symptom relief is strongest for restless legs.

• Pregnancy outcomes. The review cites large observational studies linking prenatal anemia with neurodevelopmental disorders, postpartum hemorrhage, transfusion, and severe maternal morbidity, and it notes itself that these associations may not reflect causation and that the cause of anemia was not always identified [1].

Tests to Consider

• Ferritin in any patient with anemia, microcytosis, fatigue, pica, or restless legs, and in those with heavy menstrual bleeding, pregnancy, IBD, CKD, heart failure, cancer, or bariatric surgery. Interpret below 30 ng/mL as iron deficiency regardless of the lab’s reference range; treat 30 to 50 as a gray zone in symptomatic patients.

• TSAT when inflammation is present, or when ferritin is above 30 ng/mL but suspicion remains; drawn at least 5 to 9 hours after any iron-containing food or supplement.

• Celiac serologies, H. pylori testing, and antiparietal cell antibodies when the cause is not obvious, especially with upper GI symptoms.

• Upper endoscopy and colonoscopy for men and postmenopausal women with iron-deficiency anemia and no clear source.

• Reticulocyte count about a week into treatment, and hemoglobin at about 2 weeks on oral iron.

• Serum phosphorus before a second dose of ferric carboxymaltose.

Bob’s Pearls

1. Iron deficiency precedes anemia, often by months. A normal hemoglobin and MCV do not exclude it; half of iron-deficient pregnant individuals in one study had neither anemia nor microcytosis.

2. Ferritin below 30 is iron deficiency, whatever the lab’s reference range says. It had 92% sensitivity and 98% specificity for absent marrow iron. Thirty is not universal: the AGA uses 45 in anemia, and some argue for 50. Between 30 and 50, let the patient’s symptoms and the TSAT guide you.

3. In inflammation, check TSAT. Ferritin is an acute-phase reactant; with absolute iron deficiency plus inflammation it rarely exceeds 100, and a TSAT below 20% carries the diagnosis.

4. Ask about ice and restless legs. Each occurred in about 40% of patients with iron deficiency in one large series.

5. Find the cause. In men and postmenopausal women, iron-deficiency anemia means upper endoscopy and colonoscopy; in pooled data from asymptomatic patients, they found upper GI cancer in 2% and lower GI cancer in about 9%. Test for celiac disease and gastritis when the source is unclear.

6. Every other day. Each oral dose raises hepcidin for about a day. Alternate-day dosing gives the same repletion with fewer GI side effects, but it takes twice as long.

7. IV iron is safer than its reputation. Anaphylaxis is about 1 in 200,000 infusions. Minor reactions resolve by stopping and restarting slowly; do not treat them with epinephrine and diphenhydramine. Premedication is unnecessary.

8. Transfusion is not iron therapy. It raises the hemoglobin but leaves the deficit.

References

[1] Auerbach M, DeLoughery TG, Tirnauer JS. Iron deficiency in adults: a review. JAMA. 2025;333(20):1813-1823. https://doi.org/10.1001/jama.2025.0452

[2] Tawfik YMK, Billingsley H, Bhatt AS, et al. Absolute and functional iron deficiency in the US, 2017-2020. JAMA Netw Open. 2024;7(9):e2433126. https://doi.org/10.1001/jamanetworkopen.2024.33126

[3] Achebe MO, Mandell E, Jolley K, et al. Pagophagia and restless legs syndrome are highly associated with iron deficiency and should be included in histories evaluating anemia. Am J Hematol. 2023;98(1):E8-E10. https://doi.org/10.1002/ajh.26775

[4] Avni T, Reich S, Lev N, Gafter-Gvili A. Iron supplementation for restless legs syndrome: a systematic review and meta-analysis. Eur J Intern Med. 2019;63:34-41. https://doi.org/10.1016/j.ejim.2019.02.009

[5] Chao HX, Zack T, Leavitt AD. Screening characteristics of hemoglobin and mean corpuscular volume for detection of iron deficiency in pregnancy. Obstet Gynecol. 2025;145(1):91-94. https://doi.org/10.1097/AOG.0000000000005753

[6] Mast AE, Blinder MA, Gronowski AM, Chumley C, Scott MG. Clinical utility of the soluble transferrin receptor and comparison with serum ferritin in several populations. Clin Chem. 1998;44(1):45-51. https://doi.org/10.1093/clinchem/44.1.45

[7] Martens K, DeLoughery TG. Sex, lies, and iron deficiency: a call to change ferritin reference ranges. Hematology Am Soc Hematol Educ Program. 2023;2023(1):617-621. https://doi.org/10.1182/hematology.2023000494

[8] Guyatt GH, Oxman AD, Ali M, Willan A, McIlroy W, Patterson C. Laboratory diagnosis of iron-deficiency anemia: an overview. J Gen Intern Med. 1992;7(2):145-153. https://doi.org/10.1007/BF02598003

[9] Rockey DC, Altayar O, Falck-Ytter Y, Kalmaz D. AGA technical review on gastrointestinal evaluation of iron deficiency anemia. Gastroenterology. 2020;159(3):1097-1119. https://doi.org/10.1053/j.gastro.2020.06.045

[10] Ko CW, Siddique SM, Patel A, et al. AGA clinical practice guidelines on the gastrointestinal evaluation of iron deficiency anemia. Gastroenterology. 2020;159(3):1085-1094. https://doi.org/10.1053/j.gastro.2020.06.046

[11] Mahadev S, Laszkowska M, Sundström J, et al. Prevalence of celiac disease in patients with iron deficiency anemia: a systematic review with meta-analysis. Gastroenterology. 2018;155(2):374-382.e1. https://doi.org/10.1053/j.gastro.2018.04.016

[12] Ali MD. Proton pump inhibitors’ use and risk of iron deficiency anaemia: a systematic review and meta-analysis. Curr Rev Clin Exp Pharmacol. 2023;18(2):158-166. https://doi.org/10.2174/2772432817666220307121220

[13] Stoffel NU, Cercamondi CI, Brittenham G, et al. Iron absorption from oral iron supplements given on consecutive versus alternate days and as single morning doses versus twice-daily split dosing in iron-depleted women: two open-label, randomised controlled trials. Lancet Haematol. 2017;4(11):e524-e533. https://doi.org/10.1016/S2352-3026(17)30182-5 [Author list and DOI to confirm.]

[14] von Siebenthal HK, Gessler S, Vallelian F, et al. Alternate day versus consecutive day oral iron supplementation in iron-depleted women: a randomized double-blind placebo-controlled study. EClinicalMedicine. 2023;65:102286. https://doi.org/10.1016/j.eclinm.2023.102286

[15] Okam MM, Koch TA, Tran MH. Iron supplementation, response in iron-deficiency anemia: analysis of five trials. Am J Med. 2017;130(8):991.e1-991.e8. https://doi.org/10.1016/j.amjmed.2017.03.045

[16] Avni T, Bieber A, Grossman A, Green H, Leibovici L, Gafter-Gvili A. The safety of intravenous iron preparations: systematic review and meta-analysis. Mayo Clin Proc. 2015;90(1):12-23. https://doi.org/10.1016/j.mayocp.2014.10.007

[17] Szebeni J, Fishbane S, Hedenus M, et al. Hypersensitivity to intravenous iron: classification, terminology, mechanisms and management. Br J Pharmacol. 2015;172(21):5025-5036. https://doi.org/10.1111/bph.13268

[18] Arastu AH, Elstrott BK, Martens KL, et al. Analysis of adverse events and intravenous iron infusion formulations in adults with and without prior infusion reactions. JAMA Netw Open. 2022;5(3):e224488. https://doi.org/10.1001/jamanetworkopen.2022.4488

[19] Zoller H, Wolf M, Blumenstein I, et al. Hypophosphataemia following ferric derisomaltose and ferric carboxymaltose in patients with iron deficiency anaemia due to inflammatory bowel disease (PHOSPHARE-IBD): a randomised clinical trial. Gut. 2023;72(4):644-653. https://doi.org/10.1136/gutjnl-2022-327897

[20] US Preventive Services Task Force. Screening and supplementation for iron deficiency and iron deficiency anemia during pregnancy: US Preventive Services Task Force recommendation statement. JAMA. 2024;332(11):906-913. https://doi.org/10.1001/jama.2024.15196

[21] Galetti V, Stoffel NU, Sieber C, Zeder C, Moretti D, Zimmermann MB. Threshold ferritin and hepcidin concentrations indicating early iron deficiency in young women based on upregulation of iron absorption. EClinicalMedicine. 2021;39:101052. https://doi.org/10.1016/j.eclinm.2021.101052 [Title and DOI to confirm.]

[22] Vaucher P, Druais PL, Waldvogel S, Favrat B. Effect of iron supplementation on fatigue in nonanemic menstruating women with low ferritin: a randomized controlled trial. CMAJ. 2012;184(11):1247-1254. https://doi.org/10.1503/cmaj.110950 [DOI to confirm.]

[23] Mei Z, Addo OY, Jefferds ME, Sharma AJ, Flores-Ayala RC, Brittenham GM. Physiologically based serum ferritin thresholds for iron deficiency in children and non-pregnant women: a US National Health and Nutrition Examination Surveys (NHANES) serial cross-sectional study. Lancet Haematol. 2021;8(8):e572-e582. https://doi.org/10.1016/S2352-3026(21)00168-X [DOI to confirm.]

Cirrhotic ascites is a disease of secondary hyperaldosteronism, so the aldosterone antagonist is the foundation of therapy. When urine potassium exceeds urine sodium before furosemide is added, aldosterone is still winning, and the antagonist dose should go up. The familiar spironolactone 100 mg plus furosemide 40 mg is a reasonable shortcut, but it is a convention: no trial derived it. And for the hospitalized patient with tense ascites, paracentesis comes first.

The question

Current guidance starts diuretics for cirrhotic ascites at spironolactone 100 mg and furosemide 40 mg, raising both together up to 400 mg and 160 mg [1].

That is not how I learned it in the 1970s. We started spironolactone 100 mg alone and checked a urine sodium and potassium. If urine potassium exceeded urine sodium, we increased the spironolactone, and kept increasing until sodium exceeded potassium. Only then did we add furosemide.

So did hepatology pull 100 and 40 out of the air?

The physiology

Portal hypertension lowers effective arterial blood volume, the kidney reads that as volume depletion, and aldosterone stays high. Aldosterone acts in the collecting duct, reclaiming sodium and secreting potassium in exchange. When urine potassium exceeds urine sodium, aldosterone is still winning, so increase the aldosterone antagonist. The goal is a sodium diuresis, because water follows sodium. The test reads aldosterone most cleanly before furosemide is added, because furosemide raises urine potassium on its own. That is why the old method put spironolactone first.

This explains why furosemide alone works poorly in cirrhosis. It blocks sodium reabsorption in the loop of Henle, but an aldosterone-primed collecting duct reclaims much of that sodium downstream and excretes potassium instead.

That potassium loss matters, because hypokalemia is a risk factor for hepatic encephalopathy. Low potassium stimulates ammonia production in the proximal tubule, and much of it returns to the circulation through the renal veins [2]. Gabuzda’s group showed this in cirrhosis in 1963: a diuretic lowered serum potassium, arterial ammonia rose, the extra ammonia came from the kidney, and potassium repletion reversed it. One patient developed confusion and a flapping tremor after each diuresis [3]. Avoiding hypokalemia removes one precipitant of encephalopathy, a benefit of the aldosterone antagonist beyond sodium balance. (No trial has shown that these drugs reduce encephalopathy, and they can cause hyponatremia, another precipitant.)

The evidence

The premise of my old teaching held up. In a 1983 randomized trial, Pérez-Ayuso and colleagues gave 40 nonazotemic inpatients either furosemide or spironolactone [4]. Eighteen of 19 responded to spironolactone, compared with 11 of 21 to furosemide, and 9 of the 10 furosemide failures then responded to spironolactone. The furosemide failures had the highest renin and aldosterone. Among spironolactone responders, those who needed 300 mg had markedly higher aldosterone than those who responded to 150 mg. The spironolactone dose a patient needs tracks that patient’s aldosterone. (Patients were not randomized to dose, so that comparison is observational.)

Two later trials compared starting strategies, not ratios. Santos and colleagues randomized 100 patients with moderate ascites, mostly first episodes with preserved kidney function [9], to spironolactone alone or spironolactone plus furosemide (starting at 100/40) [5]. Response rates were similar (94% vs 98%), and the combination needed dose reductions twice as often. Angeli and colleagues, in patients who mostly had recurrent ascites and often a reduced GFR [9], found that starting both drugs together resolved ascites faster and caused less hyperkalemia (4% vs 18%) [6]. That trial used potassium canrenoate, not spironolactone.

No trial derived the 100:40 ratio or compared it with other ratios. The guideline rationale is that the two drugs in that proportion leave potassium roughly neutral while moving faster than stepwise titration [8]. That is reasonable, and it is a convention.

The limits of the urine

Comparing urine sodium with potassium reads out aldosterone effect only when enough sodium reaches the collecting duct. In advanced cirrhosis with a falling GFR, the proximal tubule and loop reclaim most of the sodium, and urine sodium can stay low at any spironolactone dose.

The titration has a ceiling: 400 mg of spironolactone. Urine sodium still below potassium at that dose means diuretic-resistant ascites, which calls for paracentesis and evaluation for TIPS or transplant, not more spironolactone [1].

On the wards: tap first

Every trial above enrolled nonazotemic patients, mostly with moderate ascites. Many hospitalized patients have tense ascites, and diuretics are too slow for them. In a 1987 trial of 117 patients with tense ascites, paracentesis with albumin cleared the ascites in 97% of patients, compared with 73% on diuretics [7]. Complications during the first admission occurred in 10 patients after paracentesis and 36 on diuretics, mostly encephalopathy, kidney injury, and electrolyte disturbances, and the hospital stay was 12 days versus 31. Survival did not differ.

Paracentesis removes the fluid but not the cause. The kidney keeps retaining sodium, so most patients still need diuretics afterward; serial paracentesis is for the roughly 10% who are truly diuretic-resistant [8], and for patients who cannot tolerate diuretics because of kidney injury, hyponatremia, or encephalopathy. A practical sequence:

1. Tense ascites: large-volume paracentesis.

2. Then diuretics to prevent reaccumulation.

3. Then the urine for patients who reaccumulate. Low urine sodium on full doses means diuretic resistance; high urine sodium with weight gain means too much dietary salt.

The European guideline says to stop all diuretics for serum sodium below 125 mmol/L, acute kidney injury, worsening encephalopathy, or incapacitating muscle cramps; to stop furosemide for potassium below 3 mmol/L; and to stop the aldosterone antagonist for potassium above 6 mmol/L [9].

Tests to Consider

• Spot urine sodium and potassium in a patient who is not losing weight. Cleanest before furosemide is added; interpret cautiously on furosemide or with a reduced GFR.

• Serum potassium and creatinine within a week of starting or changing diuretics.

Bob’s Pearls

1. Cirrhotic ascites is secondary hyperaldosteronism. The aldosterone antagonist is the foundation; furosemide alone works poorly and causes hypokalemia, a risk factor for hepatic encephalopathy.

2. Before furosemide, urine potassium above urine sodium means increase the aldosterone antagonist, up to 400 mg. Beyond that, think diuretic resistance. Once furosemide is on board, the urine potassium partly reflects the furosemide.

3. The 100:40 ratio is a convention, not a discovery. It is a reasonable shortcut for keeping potassium neutral.

4. Tense ascites needs paracentesis first. Then diuretics, to keep the fluid from coming back.

References

[1] Biggins SW, Angeli P, Garcia-Tsao G, et al. Diagnosis, evaluation, and management of ascites, spontaneous bacterial peritonitis and hepatorenal syndrome: 2021 practice guidance by the American Association for the Study of Liver Diseases. Hepatology. 2021;74(2):1014-1048. https://doi.org/10.1002/hep.31884

[2] Weiner ID. Roles of renal ammonia metabolism other than in acid-base homeostasis. Pediatr Nephrol. 2017;32(6):933-942. https://doi.org/10.1007/s00467-016-3401-x

[3] Baertl JM, Sancetta SM, Gabuzda GJ. Relation of acute potassium depletion to renal ammonium metabolism in patients with cirrhosis. J Clin Invest. 1963;42(5):696-706. https://doi.org/10.1172/JCI104761

[4] Pérez-Ayuso RM, Arroyo V, Planas R, et al. Randomized comparative study of efficacy of furosemide versus spironolactone in nonazotemic cirrhosis with ascites: relationship between the diuretic response and the activity of the renin-aldosterone system. Gastroenterology. 1983;84(5 Pt 1):961-968. https://doi.org/10.1016/0016-5085(83)90198-1

[5] Santos J, Planas R, Pardo A, et al. Spironolactone alone or in combination with furosemide in the treatment of moderate ascites in nonazotemic cirrhosis: a randomized comparative study of efficacy and safety. J Hepatol. 2003;39(2):187-192. https://doi.org/10.1016/S0168-8278(03)00188-0

[6] Angeli P, Fasolato S, Mazza E, et al. Combined versus sequential diuretic treatment of ascites in non-azotaemic patients with cirrhosis: results of an open randomised clinical trial. Gut. 2010;59(1):98-104. https://doi.org/10.1136/gut.2008.176495

[7] Ginès P, Arroyo V, Quintero E, et al. Comparison of paracentesis and diuretics in the treatment of cirrhotics with tense ascites: results of a randomized study. Gastroenterology. 1987;93(2):234-241. https://doi.org/10.1016/0016-5085(87)91007-9

[8] Runyon BA. Management of adult patients with ascites due to cirrhosis: an update. Hepatology. 2009;49(6):2087-2107. https://doi.org/10.1002/hep.22853

[9] European Association for the Study of the Liver. EASL Clinical Practice Guidelines for the management of patients with decompensated cirrhosis. J Hepatol. 2018;69(2):406-460. https://doi.org/10.1016/j.jhep.2018.03.024

Bottom Line: Economists have a name for why a decision maker skips the costly work of checking whether a decision is right: a negative externality, a cost of a decision that lands on someone other than the party making it, leaving that party with no built-in reason to find out beforehand whether the decision is sound. That is the condition this series has documented four times over. A fee schedule, a set of quality measures, a guideline, and the process that writes guidelines were each decided by a body that does not bear the cost of the decision being wrong, a cost absorbed instead by physicians who must comply and patients who are treated under the result, and each became an enforceable mandate, medicolegal standard, payer compliance target, order-set template, before anyone tested whether it would do what it claimed. A new drug cannot externalize that cost onto patients the same way, because the FDA stands between the company and the prescription pad. Nothing comparable stands between a guideline panel, a quality-measure committee, or a payment-schedule body and the physician who has to comply. This post proposes a specific, inexpensive fix for that missing check: the premortem, a 90-minute structured exercise, developed by psychologist Gary Klein and tested in a handful of small studies, in which a team assumes a plan has already failed and works backward to explain why. It will not replace the outcome testing the earlier posts called for, and it will not by itself realign who bears the cost of a bad decision. It is a cheap, validated way to make the cost visible to the decision maker before the decision takes effect, rather than after. I know it works on a small scale because this series has been running an informal version of it on itself, post by post, for months.

The first three posts in this series looked at how physicians get paid, how their time is measured, and how their performance is scored. The fourth and fifth looked at guidelines: specific recommendations enforced as mandates before being tested, and a development process that lets two guideline panels look at the same evidence and reach opposite conclusions. The externality is cleanest in the first and second of those stories, a body like the RUC or a quality-measure committee is institutionally distinct from the physicians and patients who live with its decisions. It is muddier for a guideline panel, which is drawn from the same profession as the physicians who will follow its recommendations, so calling the whole series one undifferentiated externality problem would be sloppy economics. But even a guideline panel never bears the specific cost of a specific recommendation turning out to be wrong; the physician who complied and the patient who was treated do, and that is enough of a separation to blunt the same incentive to check beforehand. In each case the failure was the same shape: a system-level intervention with real consequences for patients and physicians went into effect, and enforcement, before anyone built in a mechanism to ask what could go wrong. This closing post is not another case history. It is a specific, low-cost proposal for the step that was missing each time.

A Technique With an Actual Evidence Base

In 2007, psychologist Gary Klein described a technique he called the premortem in the Harvard Business Review. [1] A postmortem asks what went wrong after a project has failed. A premortem asks the same question before the project starts, by having the team assume, as a stated fact, that it has already failed, and then write down every reason they can think of why. Klein’s own summary of the logic is direct: “a premortem may be the best way to circumvent any need for a painful postmortem.” [1]

The technique rests on a cognitive mechanism called prospective hindsight, imagining that an outcome has already occurred rather than merely considering that it might. Mitchell, Russo, and Pennington demonstrated in 1989 that this framing increases the number of reasons people generate for a future outcome by about 30 percent, compared with simply being asked to consider that the outcome might occur. [2] I want to be precise about what that finding does and doesn’t show, since the series has been consistent about not overstating what a cited study found. Mitchell and colleagues measured the quantity of reasons generated, not their quality or accuracy, and even Klein’s own 2007 description of the finding rounds it up slightly, describing it as increasing people’s “ability to correctly identify reasons,” a stronger claim than the original study tested. [1,2] The honest version is narrower: imagining a failure as a settled fact makes people generate more candidate explanations. Whether those explanations are the right ones is a separate question, one later studies tried to address directly.

A 2010 study by Veinott, Klein, and Wiggins asked 178 university students, working in small groups, to evaluate a hypothetical H1N1 lockdown plan under one of five conditions: no evaluation, a two-minute open critique, a premortem, a pro/con list, or a cons-only list. [3] The premortem produced the largest drop in each group’s confidence in the plan, an average fall of 25 points on their rating scale, compared with 14 points for the pro/con condition, a statistically significant difference. [3] A drop in confidence is not the same as a better plan, but it is evidence that the premortem’s framing surfaces doubts that a straightforward pro/con exercise does not.

A 2016 study by Gallop and Bischoff came closer to testing output quality directly. [4] They randomly assigned 101 experienced program managers and engineers to teams trained in either traditional brainstorming or the premortem technique, then gave each team an identical scenario (adding a new fuse mechanism to an existing weapons system) and 45 minutes to identify risks and recommend changes. An expert panel, blinded to which technique each team had used, scored the risks each team generated for creativity and plausibility. Premortem teams identified significantly more risks that met the panel’s bar for both creative and plausible, and did better specifically at identifying the “black swan” risks, the ones that had actually caused a comparable real-world failure in the historical case the scenario was based on, but that a conventional brainstorm tends to miss. [4] The authors attributed the difference to the premortem’s framing neutralizing what they called deliberate ignorance, the tendency of a team member who senses a problem to stay quiet rather than be the one who sounds pessimistic about a plan everyone else is already invested in.

None of these are large or definitive studies, and I am naming that limitation deliberately, because a series built on asking whether an intervention has been tested should not then propose a fix without disclosing how thin that fix’s own evidence base is. Across three studies, the premortem has been shown to generate more candidate explanations, reduce overconfidence more than a simple pros-and-cons list, and surface risks a conventional brainstorm misses. It has not been tested, as far as I can find, against the actual outcome that matters here: whether teams that run a premortem before a decision make better decisions, measured by what happens later. That is a real gap, and it is worth sitting with for a moment, because it is precisely the kind of gap this series has spent four posts criticizing medicine for tolerating in its own interventions.

Where This Series Already Ran One

Here is where I can offer something closer to a worked example than a published study. Every post in this series, including the one you are reading, has gone through a version of this exercise before publication: a deliberate, adversarial pass in which I ask what is wrong with the draft, as though its errors had already happened and I were explaining them after the fact, rather than simply reviewing it once more for polish. That is a premortem, run on a blog post, in about the time it takes to read this section.

It caught real mistakes. An early draft of the fourth post in this series framed the CAST antiarrhythmic trial as though it had been driven by a formal clinical practice guideline; the premortem-style pass asked what a skeptical reader would attack first, and the answer was that no such guideline existed in the mid-1980s, which led to reframing that section as a pre-guideline-era warning instead. The same post briefly described a hospital as “penalized” for a compliance flag that the body of the text only established as a noncompliance flag, an overclaim a close rereading caught before publication. A later full rereading of that post caught an arithmetic error that had crept in silently: an early draft’s running count of “five cases” no longer matched the actual number of cases in the post, after a case had been added during revision without the count being updated anywhere else. And the fifth post initially used a gout-guideline dispute as its sharpest example of specialty-driven conflict of interest, until that same adversarial pass noted that the best available account of the dispute attributed much of it to an evidence-timing gap rather than panel bias, which meant the example was quietly undermining its own argument and needed to come out.

None of those are catastrophic errors. That is the point. A premortem does not require the draft to be bad, or the guideline panel to be dishonest, or the quality measure committee to be careless, to be worth running. It is useful precisely because competent people produce work with exactly this kind of recoverable flaw, a misattributed cause, an overclaimed word, a stale number, a case that undercuts its own point, and those flaws are cheap to catch before publication and expensive to catch after.

What a Premortem Would Have Asked

Apply the same exercise, run before enforcement rather than after a draft, to three cases this series has already documented.

The 1992 American College of Physicians guideline on hormone therapy recommended that women with or at risk for coronary heart disease were likely to benefit from estrogen plus progestin, based on observational data, and said so explicitly: the guideline’s own text stated that randomized trials would be needed to confirm the effect and that the recommendation could change once they were run. [5] A premortem does not require predicting that the Women’s Health Initiative would later find increased cardiovascular and breast cancer risk; the panel had already told its readers, in print, that this was possible. [6] What a premortem would have forced is a second question the guideline never answered: given that we know this recommendation might be reversed, what are we doing in the meantime to make sure it gets reviewed on a schedule rather than left as an open-ended mandate? That recommendation functioned as a guideline, and therefore a standard of care, for roughly a decade before the trial that tested it reported. A premortem’s output here would not have been a different recommendation. It would have been a built-in trigger to revisit it.

The 2009 vancomycin trough target of 15 to 20 mg/L was, by the guideline authors’ own later admission, derived largely from retrospective, single-center, observational studies, adopted as a practical bedside stand-in for a pharmacodynamic target that was hard to calculate at the time. [7] A premortem run before that target went into hospital order sets nationwide would have asked the question the panel was well positioned to answer honestly: if this number from a handful of single-center studies turns out to be wrong, how would we know, and on what schedule would we check? The actual answer was eleven years, during which the target hardened into standard practice before a 2020 revision replaced it with an AUC-based target that the same authors acknowledged carried similar evidentiary limits. [8] A scheduled review is a cheap thing to build into a guideline at the moment it is written. It is not something that gets added after the fact, because by then there is no longer a mechanism prompting anyone to ask.

The four-hour pneumonia rule and SEP-1 both became enforceable, publicly reported quality measures on the premise that specifying a time window for antibiotics would improve outcomes. In both cases, the harm, misdiagnosis and inappropriate antibiotic use for the pneumonia rule [9,10], overdiagnosis and antibiotic overuse without a demonstrated mortality benefit for SEP-1 [11], was documented only after each measure was already in wide national use, by researchers and specialty societies working backward from real consequences. A premortem run before either measure was finalized would have asked the exact question those later papers ended up answering empirically: assume, five years from now, that this measure has increased unnecessary antibiotic use without improving the outcome it was meant to protect. Why would that have happened? Clinicians who treat patients who do not actually have the target condition, under time pressure, in ambiguous cases, were available to answer that question in the room, before the measure took effect, not years into an infectious disease society’s position paper after it.

In each case, the information a premortem would have surfaced was not hidden. The guideline panel that wrote the HRT recommendation stated the limitation in the guideline itself. The vancomycin panel’s own later paper acknowledged the original evidence base’s limits. The physicians who eventually documented harm from the pneumonia and sepsis measures were practicing the whole time the measures were in effect, available to be asked. What was missing was not the knowledge. It was a scheduled moment, built into the process before the mandate took effect, when someone was required to ask the question and someone else was required to listen.

Not a Substitute, a First Gate

To be clear about what this proposal is not: it is not an argument that a premortem replaces the outcome testing the earlier posts in this series called for, and it does not resolve the conflicts of interest or the conflicting-guidelines problem documented in the fifth post in this series. A measure still needs to be tested against the outcome it claims to improve. A guideline still needs to disclose and manage financial and intellectual conflicts of interest on its panel. Two professional societies that disagree still need some mechanism for reconciling, or at least flagging, that disagreement to the physicians caught between them. A 90-minute exercise does none of that.

What it does is sit in front of all of it, cheaply, at the one moment in the process when raising a doubt is still inexpensive: before the recommendation, the measure, or the fee schedule takes effect and starts functioning as a mandate. Spellberg’s account of the aztreonam guideline clause from the fourth post in this series makes the contrast concrete. The committee ran out of time at a meeting before it could discuss that part of the recommendation table, so, in the guideline’s first author’s own words, “it never got vetted.” [12] A premortem is structurally built to prevent exactly that failure, because the step is scheduled in advance, with time protected for it, rather than squeezed into whatever is left at the end of an agenda. It does not require a new institution, a new federal rule, or a new testing infrastructure. It requires a guideline panel, a quality-measure committee, or a payment-schedule body to set aside 90 minutes, before finalizing a recommendation, to assume it has already failed and write down why.

Bob’s Pearls

1. A negative externality is a cost of a decision that lands on someone other than the party making it. A fee schedule, a set of quality measures, and a guideline are each decided by a body that does not bear the cost of being wrong, which is exactly the condition that removes any built-in reason to check beforehand.

2. A new drug gets controlled testing and post-marketing surveillance because regulators assume it can cause harm. None of the four interventions documented in this series, a fee schedule, quality measures, a guideline, or the guideline development process, got anything comparable before taking effect.

3. The premortem, developed by Gary Klein in 2007, has a real but modest evidence base: it increases the number of candidate explanations people generate for a hypothetical failure by about 30 percent, though the original study measured quantity of reasons, not their accuracy.

4. In a 2010 study, premortems produced a larger drop in overconfidence (25 points) than a pro/con list (14 points) among 178 participants evaluating a hypothetical plan.

5. In a 2016 study, experienced program managers trained in the premortem technique identified more creative and plausible risks, including the specific “black swan” risks a comparable real failure had actually involved, than teams using conventional brainstorming.

6. The premortem’s own evidence base has a real limitation worth naming honestly: no study I could find tested whether premortems improve the real-world outcomes of the decisions they’re used on, only whether they improve risk identification and reduce overconfidence in the room.

7. This series’ own sparring-partner critique passes, an informal premortem run on each draft before publication, caught real errors: a misattributed guideline-driven framing for the CAST trial, an overclaimed word (”penalized”), a silent arithmetic error, and a case study that undermined its own argument.

8. In each documented failure, the information a premortem would have surfaced already existed, in the guideline’s own text, in a later author’s own admission, or in practicing clinicians’ daily experience. What was missing was a scheduled moment to ask the question before the mandate took effect, not the knowledge to answer it.

References

1. Klein G. Performing a project premortem. Harv Bus Rev. 2007;85(9):18-19.

2. Mitchell DJ, Russo JE, Pennington N. Back to the future: Temporal perspective in the explanation of events. J Behav Decis Mak. 1989;2(1):25-38. doi:10.1002/bdm.3960020103

3. Veinott B, Klein G, Wiggins S. Evaluating the effectiveness of the PreMortem technique on plan confidence. Proceedings of the 7th International ISCRAM Conference. Seattle, WA; 2010:1-9.

4. Gallop A, Bischoff J. How to catch a black swan: measuring the benefits of the premortem technique for risk identification. J Enterp Transform. 2016;6(2). doi:10.1080/19488289.2016.1240118

5. American College of Physicians. Guidelines for counseling postmenopausal women about preventive hormone therapy. Ann Intern Med. 1992;117(12):1038-1041. doi:10.7326/0003-4819-117-12-1038

6. Writing Group for the Women’s Health Initiative Investigators. Risks and benefits of estrogen plus progestin in healthy postmenopausal women: principal results from the Women’s Health Initiative randomized controlled trial. JAMA. 2002;288(3):321-333. doi:10.1001/jama.288.3.321

7. Rybak MJ, Lomaestro BM, Rotschafer JC, et al. Therapeutic monitoring of vancomycin in adult patients: a consensus review of the American Society of Health-System Pharmacists, the Infectious Diseases Society of America, and the Society of Infectious Diseases Pharmacists. Am J Health Syst Pharm. 2009;66(1):82-98. doi:10.2146/ajhp080434

8. Rybak MJ, Le J, Lodise TP, et al. Therapeutic monitoring of vancomycin for serious methicillin-resistant Staphylococcus aureus infections: a revised consensus guideline and review. Am J Health Syst Pharm. 2020;77(11):835-864. doi:10.1093/ajhp/zxaa036

9. Kanwar M, et al. Misdiagnosis of community-acquired pneumonia and inappropriate utilization of antibiotics: side effects of the 4-h antibiotic administration rule. Chest. 2007;131:1865-1869.

10. Wachter RM, Flanders SA, Fee C, Pronovost PJ. Public reporting of antibiotic timing in patients with pneumonia: lessons from a flawed performance measure. Ann Intern Med. 2008;149(1):29-32. doi:10.7326/0003-4819-149-1-200807010-00007

11. Ford JS, Morrison JC, Kyaw M, et al. The effect of Severe Sepsis and Septic Shock Management Bundle (SEP-1) compliance and implementation on mortality among patients with sepsis: a systematic review. Ann Intern Med. 2025;178(4):543-557. doi:10.7326/ANNALS-24-02426

12. Centor RM, Spellberg B, Shaneyfelt T. Annals On Call: Guidelines With Humility [podcast]. Ann Intern Med. doi:10.7326/A20-0026

Bottom Line: The previous post in this series documented six cases in which a specific guideline recommendation was enforced as a mandate before it had been tested, and in most cases was later reversed. This post is about a related but distinct failure: the process that produces guidelines in the first place, and what happens when that process runs twice on the same question and comes back with two different answers. Guideline panels are supposed to meet explicit standards for managing conflicts of interest and grading evidence. In practice, financial conflicts of interest remain common among panel members, non-financial conflicts (a panelist’s specialty, training, or research investment shaping which evidence feels compelling) are harder to measure and rarely disclosed at all, and compliance with the field’s own trustworthiness standards is inconsistent. The result is predictable: guidelines from different professional societies, drawing on the same underlying evidence, sometimes tell physicians to do opposite things. A mandate built on untested evidence is one problem. Two mandates that contradict each other, both enforced, is a second one, and it is the subject of this post.

The first three posts in this series looked at how physicians get paid, how their time is measured, and how their performance is scored. The fourth looked at specific guideline recommendations that were adopted and enforced before being tested against the outcomes they were meant to improve. This post asks a different question about the same machinery: how trustworthy is the process that writes guidelines in the first place, and what happens when two groups of experts, working from the same evidence, publish guidelines that disagree?

What a Trustworthy Guideline Is Supposed to Look Like

In 2011, the Institute of Medicine (now the National Academy of Medicine) published a report, Clinical Practice Guidelines We Can Trust, laying out explicit standards for how a guideline should be developed: a systematic, transparent review of the evidence; a multidisciplinary panel; explicit management of financial and intellectual conflicts of interest; external review; and clear, graded links between the strength of the evidence and the strength of each recommendation. [1] The report existed because, by 2011, there was already substantial evidence that guidelines routinely fell short of those standards. A 1999 analysis in JAMA evaluated 279 guidelines produced by 69 different organizations between 1985 and 1997 against a validated quality-assessment instrument. Overall methodological quality, scored on a 0 to 100 scale, improved from a mean of 36.9 in the earliest period studied to 50.4 in the most recent, still well short of what the instrument’s own benchmarks would call adequate. [2]

A later audit suggests that gap did not close on its own. A 2012 review in Archives of Internal Medicine scored 114 guidelines randomly sampled from the National Guideline Clearinghouse against the IOM’s own trustworthiness standards and found that the typical guideline met 44 percent of them. [3] Guidelines are, collectively, graded against a standard the field itself wrote, and the field’s own output does not consistently meet it.

Who Writes Them

The IOM standards call for managing financial conflicts of interest on guideline panels, ideally by excluding members with significant ones from voting on related recommendations. A 2011 study in BMJ examined disclosed and undisclosed financial relationships among 288 panel members who had participated in developing U.S. and Canadian guidelines on hyperlipidemia and diabetes screening and treatment. Fifty-two percent had a financial conflict of interest with industry, and of those, nearly one in ten had failed to disclose it even when they formally stated they had none. Guidelines sponsored by government bodies had far lower rates of panelist conflict (16 percent) than those sponsored by non-government organizations (69 percent). [4] Disclosure, even accurate disclosure, is not the same as exclusion from the vote, and the study did not find a consistent process across these guidelines for managing a conflict once it was disclosed. The 2012 trustworthiness audit cited above found the same pattern at the top of the panel: more than 70 percent of guideline chairs and more than 90 percent of co-chairs had a financial conflict of interest, even though the IOM standards the audit was measuring against specifically recommend that panel leadership be conflict-free. [3]

Financial conflicts are only half of this problem that shows up on a disclosure form. On a 2021 Annals on Call podcast episode, I asked Terry Shaneyfelt, the author of the 1999 methodological-quality study cited above and a veteran of multiple guideline panels, about the conflicts that never make it onto that form. “There are intellectual conflicts of interest, and this can apply to the specialty that you’re in,” he said. “Somebody who’s a urologist might think about approaching prostate cancer just differently than a primary care physician might think about approaching it.” A second layer sits on top of that one: “Intellectual conflicts at another level that might be related to the research that you do on a particular topic, that you really are so invested so heavily in a knowledge base... that you’re conflicted and maybe not open to seeing some alternative views.” [10] Neither conflict requires bad faith. A cardiologist and a primary care physician asked to weigh the same observational signal about a cardiac drug do not start from the same place, not because either is dishonest, but because years of training and research investment have shaped what each one finds intuitively convincing. This kind of conflict is harder to name, harder to disclose, and essentially absent from the trustworthiness audits described above, which can count financial relationships but have no comparable instrument for counting this.

When Two Guidelines Disagree

Even a guideline developed in reasonably good faith, by a panel managing its financial conflicts responsibly, can collide with another guideline written by a different panel looking at the same evidence. In a 2017 editorial in Annals of Internal Medicine, Sheldon Greenfield and Sherrie Kaplan laid out why this keeps happening and why it matters well beyond academic disagreement: guidelines are now tied directly to physician compensation and quality measurement, so when two guidelines on the same question disagree, physicians are being held accountable to two different, and sometimes incompatible, standards of care at once. [5]

Their clearest example is breast cancer screening. The American College of Obstetricians and Gynecologists and the American Cancer Society have recommended mammography starting at age 40; the U.S. Preventive Services Task Force has recommended starting at 50, citing concern about the cumulative harms of false positives and overdiagnosis in the intervening decade. A 2017 survey by Radhakrishnan and colleagues found that primary care physicians rated the ACOG and ACS guidelines as the most trustworthy of the three, the same guidelines whose approach, by the USPSTF’s own analysis, carries the greatest risk of harm from overscreening. [5] Trust in a guideline and the guideline’s actual risk-benefit balance are not the same thing, and physicians have no reliable way to tell the two apart from outside the panel room.

Greenfield and Kaplan’s second example is glycemic targets in type 2 diabetes, where the VA/DoD, the American Diabetes Association, and the American Association of Clinical Endocrinologists have each published different target A1c ranges for different patient subgroups, driven in part by genuine disagreement about how to define “high risk” and in part, they argue, by the same financial and intellectual conflicts of interest documented above. [5] A physician managing diabetes is not choosing among equally valid clinical judgments here; she is choosing which of three professional societies’ mandates to be out of compliance with.

A similar pattern shows up in gout, and I spent a full podcast episode walking through it with Shaneyfelt. In 2017, the American College of Physicians published a guideline stating there was no trial evidence that titrating allopurinol to a specific uric acid target below 6 mg/dL was any better than simply raising the dose until flares stopped. [6] The same year, the Gout, Hyperuricemia and Crystal-Associated Disease Network, an international group of rheumatologists, published a consensus statement in Nature Reviews Rheumatology arguing the opposite, citing physiologic evidence that lowering uric acid below that threshold reduces flares. [7] Shaneyfelt’s own read of the dispute, when I asked him about it, was more charitable to both sides than a simple specialty-bias story: part of the rheumatology guideline’s supporting evidence was published after the ACP guideline had already gone to press, and “guidelines should undergo regularly updating... most of them are outdated for sure by five years.” [10] But he did not let the deeper problem off the hook. As he put it, “the funny thing is at the very top of the evidence hierarchy is a guideline, and many guidelines, which is supposed to be the peak of evidence translation, relies on something that’s at the very bottom of the evidence pyramid.” His proposed fix is blunt: when a guideline panel is working from expert opinion or physiologic reasoning rather than trial evidence, “you need to be very explicit in saying this is expert opinion... this is not a guideline,” and it should be labeled a consensus statement instead, precisely so that no payer or regulator mistakes it for a tested recommendation and converts it into a quality measure. [10]

International comparisons make the same point with less of the discipline-versus-discipline framing. A 2007 review by Matthys and colleagues compared national guidelines for managing sore throat across multiple European countries and found substantial disagreement on basic questions: when to test for group A streptococcus, when to prescribe antibiotics at all, and which antibiotic to use first. A 2021 follow-up by Coutinho and colleagues, comparing sore-throat guidelines from 24 countries, found the disagreement had not resolved. [8,9] The guidelines in both reviews were citing much of the same underlying trial literature. What differed was which studies each panel chose to weight most heavily, a choice that is not fully specified by the evidence itself and leaves room for exactly the kind of institutional and specialty-driven preference described above. Sore throat is a useful test case precisely because the stakes are low and the evidence base is old and relatively stable; if guideline panels can disagree this much here, the mechanism producing that disagreement is unlikely to be limited to higher-stakes, more contested conditions.

What This Means for the Mandate Mechanism

The previous post in this series argued that a guideline functions as a mandate whether or not its authors intend it to: a medicolegal standard, a payer and regulatory compliance target, the template for a hospital order set. Put that mechanism together with what this post has documented, inconsistent compliance with the field’s own trustworthiness standards, financial conflicts of interest in roughly half of one studied panel, an entire category of non-financial, specialty-driven conflict that no audit currently measures, and two or three professional societies that can look at the same evidence and mandate opposite practices, and the shape of the problem becomes clearer. The third post in this series showed that quality measures are built on guidelines without separately testing whether following the measure improves outcomes. When the underlying guideline a measure is built on is itself one of two or three competing, conflicting guidelines, the measure is not just untested; it is, for an unknown fraction of the physicians and patients it is applied to, actively enforcing the wrong answer to a question that genuinely has more than one.

None of this is an argument against guidelines. It is an argument that the current process, for writing them, for disclosing and managing conflicts of interest, and for reconciling disagreement between them, does not yet meet the standard of rigor the mandate mechanism requires. Shaneyfelt’s own suggested fix, reserving the word “guideline” for recommendations built on a real evidence base and labeling the rest honestly as consensus statements or expert opinion, is a modest, achievable version of that rigor, and it would already prevent some of the cases in the previous post. The final post in this series turns to a more complete proposal for closing that gap before a guideline, rather than after one.

Bob’s Pearls

1. A guideline is graded against trustworthiness standards the field itself wrote in 2011. A 2012 audit of 114 guidelines sampled from the National Guideline Clearinghouse found the typical guideline met only 44 percent of those standards.

2. Guideline quality has improved over time but started from a low baseline and remains well short of the field’s own benchmark: a 1999 analysis found mean quality scores rising from 36.9 to 50.4 out of 100 across guidelines published 1985-1997.

3. Financial conflicts of interest on guideline panels are common, and under-reported even when they’re common: 52 percent of panelists in one study had a financial conflict of interest, nearly 1 in 10 of them undisclosed, and more than 70 percent of panel chairs and 90 percent of co-chairs had one, despite IOM standards calling for conflict-free leadership.

4. A second kind of conflict, a panelist’s own subspecialty, training, and research investment shaping which evidence feels persuasive, is essentially unmeasured by current disclosure and audit instruments, even though it may matter as much as the financial kind.

5. Mammography screening guidelines from different societies recommend starting at different ages; the guideline physicians rate as most trustworthy is, by the evidence, the one with the greatest overdiagnosis risk.

6. Diabetes, gout, and sore-throat guidelines all show the same pattern: different professional societies, looking at much of the same evidence, reach different and sometimes opposite recommendations. The reasons vary, sometimes a genuine evidence-timing gap, sometimes specialty-driven preference, but the effect on the physician caught between two mandates is the same either way.

7. When a quality measure is built on one of two or more conflicting guidelines, the measure can be actively enforcing the wrong answer for a meaningful share of the patients it applies to.

8. Guidelines sit, by definition, at the top of the evidence hierarchy, yet many of them rest on evidence from the bottom of it, expert opinion or physiologic reasoning. A guideline built that way should say so plainly rather than being published, and converted into a quality measure, as if it were something else.

References

1. Institute of Medicine. Clinical Practice Guidelines We Can Trust. Washington, DC: National Academies Press; 2011.

2. Shaneyfelt TM, Mayo-Smith MF, Rothwangl J. Are guidelines following guidelines? The methodological quality of clinical practice guidelines in the peer-reviewed medical literature. JAMA. 1999;281(20):1900-1905. doi:10.1001/jama.281.20.1900

3. Kung J, Miller RR, Mackowiak PA. Failure of clinical practice guidelines to meet Institute of Medicine standards: two more decades of little, if any, progress. Arch Intern Med. 2012;172(21):1628-1633. doi:10.1001/2013.jamainternmed.56

4. Neuman J, Korenstein D, Ross JS, Keyhani S. Prevalence of financial conflicts of interest among panel members producing clinical practice guidelines in Canada and United States: cross sectional study. BMJ. 2011;343:d5621. doi:10.1136/bmj.d5621

5. Greenfield S, Kaplan SH. When clinical practice guidelines collide: finding a way forward [Editorial]. Ann Intern Med. 2017;167(9):677-678. doi:10.7326/M17-2665

6. Qaseem A, Harris RP, Forciea MA; Clinical Guidelines Committee of the American College of Physicians. Management of acute and recurrent gout: a clinical practice guideline from the American College of Physicians. Ann Intern Med. 2017;166(1):58-68. doi:10.7326/M16-0570

7. Dalbeth N, Bardin T, Doherty M, et al. Discordant American College of Physicians and international rheumatology guidelines for gout management: consensus statement of the Gout, Hyperuricemia and Crystal-Associated Disease Network (G-CAN). Nat Rev Rheumatol. 2017;13(9):561-568. doi:10.1038/nrrheum.2017.126

8. Matthys J, De Meyere M, van Driel ML, De Sutter A. Differences among international pharyngitis guidelines: not just academic. Ann Fam Med. 2007;5(5):436-443. doi:10.1370/afm.741

9. Coutinho G, Duerden M, Sessa A, Caretta-Barradas S, Altiner A. Worldwide comparison of treatment guidelines for sore throat. Int J Clin Pract. 2021;75:e13879. doi:10.1111/ijcp.13879

10. Centor RM, Shaneyfelt TM. Annals On Call: The Problem of Conflicting Guidelines [podcast]. Ann Intern Med. 2021. doi:10.7326/A20-0012

After more than 45 years of practicing medicine, I’m confident I have missed Wernicke’s encephalopathy. I can’t tell you how many times, and I can’t name the patients. That is the nature of this diagnosis. A missed case of Wernicke’s doesn’t announce itself. The patient gets labeled confused, depressed, demented, or difficult, and the story ends there.

Recently we treated a patient who made me think about all of this.

The case

A man was admitted after a period of eating very little. A high-dose loop diuretic had been started shortly before. On an earlier hospitalization, he had trouble concentrating.

This time he wanted to be left alone and refused nearly every request. He became angry and threatened staff who tried to examine him.

For his first 24 hours in the hospital, he refused to eat. We started high-dose IV thiamine. Within a day he was cooperative and appeared to understand his situation. Within two days he was a new man: engaged, interested in his health, and very hungry.

Did he have Wernicke’s?

Probably. But I have to be honest about what we know.

We never documented nystagmus or tested his gait; you can’t examine the eyes of a man threatening the people trying to examine him. We didn’t get an MRI or a thiamine level. So the diagnosis rests mainly on how he responded to treatment.

Some of the usual alternative explanations don’t fit well. He refused food for his first day in the hospital, so eating didn’t drive his recovery; his appetite came back after the thiamine, not before. Still, thiamine wasn’t the only thing that changed. His medications were reviewed, his fluids and electrolytes were attended to, and he had time, and hospital delirium sometimes clears on its own.

I believe the timing, speed, and completeness of his turnaround point to thiamine. A skeptic could reasonably disagree.

Here is the uncomfortable part. Giving thiamine empirically was the right call, because waiting for certainty risks permanent brain damage. But treating early also means we rarely find out whether we were right. I can’t count my misses, and I can’t fully count my hits either.

When should we think of Wernicke’s?

The classic triad of confusion, eye findings, and ataxia is the exception. In Harper’s necropsy series, only about 16% had all three. A change in mental status is the most consistent finding, and it takes several forms:

• Apathy, inattention, and withdrawal, the most common

• Irritability, agitation, and aggression, which pull us toward a psychiatric or behavioral label

• Drowsiness progressing to stupor or coma

The setting matters as much as the signs. Risk rises with:

• Weeks of poor intake or significant weight loss

• Persistent vomiting, including hyperemesis gravidarum

• Bariatric or other GI surgery

• Cancer, especially with chemotherapy or poor intake

• Dialysis

• Alcohol use disorder, still the most common setting

• Refeeding after starvation, and dextrose given without thiamine

• Possibly loop diuretics and GLP-1 receptor agonists (see below)

My threshold: this list covers a lot of hospitalized patients, and I’m comfortable with that. When a patient has one of these risk factors and an unexplained change in mental status, of any kind, I would give thiamine. It is cheap, serious reactions are rare, and the cost of missing Wernicke’s is permanent memory loss or death. I’m not advocating thiamine for everyone, only a much lower threshold than most of us were taught.

Loop diuretics: a plausible contributor

Most of us were never taught that diuretics can deplete thiamine, and the evidence is suggestive rather than settled. In Seligmann’s 1991 pilot study, 21 of 23 heart failure patients on long-term furosemide had biochemical evidence of thiamine deficiency, compared with 2 of 16 controls. But those patients took 80 to 240 mg daily for months, and the study measured a lab marker, not clinical Wernicke’s. In healthy volunteers, urinary thiamine excretion rose with IV furosemide and returned to baseline after it was stopped.

Published cases of Wernicke’s attributed to diuretics are uncommon, and our patient had weeks of exposure rather than months. But his dose was high, and urinary thiamine loss tracks urine flow. For a patient already eating poorly, a new high-dose diuretic is one more reason to lower the threshold.

GLP-1 receptor agonists: a signal worth knowing

These drugs suppress appetite, and nausea and vomiting are their most common side effects. That is the same physiology that causes Wernicke’s after bariatric surgery and in hyperemesis gravidarum.

Case reports are accumulating:

• A pharmacovigilance study of the FDA’s adverse event database and the literature found 15 cases linked to GLP-1 drugs, mostly semaglutide or tirzepatide, and almost all reported in 2023–2024. Most patients had weight loss, vomiting, poor appetite, or malnutrition, and only 2 had the full triad.

• A French review collected 18 more cases. Nausea, vomiting, or reduced intake was present in 68%, with weight loss of 3.5 to 13.3 kg per month over 3 to 6 months.

These numbers need context. Adverse event databases can’t give an incidence rate, and reporting tends to rise once an association gets media attention, which this one has. Against millions of users, the absolute risk appears small. But many prescribers never think of thiamine, and there is no unified guidance yet on monitoring nutritional deficiencies during rapid weight loss on these drugs. A patient on a GLP-1 drug who is vomiting, losing weight quickly, and becoming confused or withdrawn deserves thiamine.

How often do we miss it?

The best evidence comes from pathologists, and it is striking:

• In Clive Harper’s Perth series, 51 cases were found at necropsy over four years, about 1.7% of autopsies, and only seven had been diagnosed during life. In his larger analysis of 131 necropsy-proven cases, 80% had not been diagnosed during life.

• In a Brazilian series, 11 of 36 autopsied patients with Wernicke’s had become comatose, and none had been diagnosed in life.

• Missed diagnosis is worse in patients without alcohol use disorder: roughly 68% undiagnosed with alcoholism versus up to 94% without.

These studies have real limits. Harper’s patients came from Perth in the 1970s and 1980s, a population heavily affected by alcohol, and included coroner cases. Australia later fortified bread with thiamine, partly because of his work. Many lesions found at autopsy are also old. In a Norwegian series of 279 consecutive autopsies, three of the four Wernicke’s cases were chronic and only one was active, and just one had been suspected clinically. An old lesion found at autopsy is not the same as a treatable episode missed during the patient’s final illness.

Delay does matter. In a Spanish series of non-alcoholic patients, everyone diagnosed within 18 days had only mild sequelae, and the authors concluded that delayed diagnosis worsens the prognosis.

What we can’t know

Almost all of this evidence comes from autopsies, and autopsies are now rare. We have no modern estimate of how often we miss Wernicke’s. It may be less often than in Harper’s era: MRI is widely available, and routine thiamine for patients with alcohol use disorder is standard. It may not be. Patients without alcohol use disorder, the group we miss most, now include people after bariatric surgery and those on GLP-1 drugs. Some missed patients die, some live on with memory loss attributed to dementia, and some improve without anyone knowing why.

What I would tell my younger self

You will not be able to prove most of your Wernicke’s diagnoses, and you will never know about the ones you missed. Act anyway. When a malnourished patient is confused, withdrawn, or angry, give thiamine first and keep thinking about the diagnosis. A rapid response supports it but doesn’t prove it. A missed case may never be counted, but the patient lives with the consequences.

References

1. Harper CG, Giles M, Finlay-Jones R. Clinical signs in the Wernicke-Korsakoff complex: a retrospective analysis of 131 cases diagnosed at necropsy. J Neurol Neurosurg Psychiatry. 1986;49:341–345.

2. Harper C. Wernicke’s encephalopathy: a more common disease than realised. A neuropathological study of 51 cases. J Neurol Neurosurg Psychiatry. 1979;42:226–231.

3. Lana-Peixoto MA, Dos Santos EC, Pittella JE. Coma and death in unrecognized Wernicke’s encephalopathy: an autopsy study. Arq Neuropsiquiatr. 1992;50:329–333.

4. Vege A, et al. Wernicke’s encephalopathy in an autopsy material obtained over a one-year period. APMIS. 1991.

5. Spanish non-alcoholic Wernicke’s series. Neurología. 2011. PubMed 21565430.

6. Seligmann H, Halkin H, Rauchfleisch S, et al. Thiamine deficiency in patients with congestive heart failure receiving long-term furosemide therapy: a pilot study. Am J Med. 1991;91:151–155.

7. Glucagon-like peptide-1 receptor agonists and Wernicke encephalopathy: a pharmacovigilance study and literature review. Clin Nutr. PubMed 41534460.

8. Gras C, De Wit V, Oussedik N, et al. Semaglutide-induced Wernicke encephalopathy: a comprehensive analysis. Eur J Clin Nutr. 2025;79:1160–1163.

Bottom Line: A clinical practice guideline is not merely a suggestion. Once published, it functions as a mandate: it becomes a medicolegal standard of care that drives lawsuits, a compliance target for hospital regulators and payers, and the template for institutional order sets, whether or not its authors intended that. A new drug cannot reach a patient until a trial has tested whether it does what it claims. A guideline recommendation can reach every physician in the country, and compel their practice through exactly these mechanisms, on the strength of a small trial, an observational association, or a plausible mechanism, with no requirement that the recommendation itself be tested before it becomes enforceable. The six cases in this post, one an anecdote and five well-documented reversals, show what that mandate mechanism looks like in practice: a specific recommendation, built on weak evidence, enforced as though it were settled, and in most cases pulled back only after real or clearly documented harm; in one case, only after a careful test showed the picture was more complicated than either side expected. A separate, related problem, the guideline development process itself, who writes guidelines, how conflicted they are, and what happens when two guidelines on the same question disagree, is the subject of the next post in this series.

The first three posts in this series looked at how physicians get paid, how their time is measured, and how their performance is scored. This post turns to the recommendations that tell them what to do in the first place, and to a fact that is easy to miss from inside the system: a guideline does not stay a suggestion. In a 2021 Annals of Internal Medicine essay, my co-authors Brad Spellberg, William Wright, and Terry Shaneyfelt and I made this point directly: “Once guidelines are published, they unfortunately tend to be treated ‘as if stone tablets descended from the heavens.’” They set medicolegal standards of care that drive lawsuits, government and private payer standards, subspecialty billing practices, and hospital order sets and regulatory compliance standards. As a result, regardless of whether they are intended to do so, guidelines functionally mandate recommended care. [1] On the podcast episode built around that essay, Brad put the same point more bluntly: guideline authors typically write language saying a recommendation isn’t intended to set a standard of care, but it does anyway. [2] A guideline is where “sexy new idea” and “not yet adequately tested” most often meet, because it can be built from a compelling mechanism or a single well-designed but small trial, and then enforced on millions of patients before a larger trial has a chance to confirm or contradict it.

A Warning From Before Guidelines Existed

Long before clinical practice guidelines existed in anything like their current form, medicine already had a clear demonstration of what happens when a plausible-sounding assumption gets turned into standard practice without being tested against an actual outcome. I’m including it first for that reason: it is close to the best warning the field already had on hand, and the guideline movement that followed built guidelines anyway without fully absorbing the lesson.

Patients who develop frequent premature ventricular contractions (PVCs) after a heart attack have a higher risk of sudden death than patients who do not. Encainide and flecainide were FDA-approved in the mid-1980s based on their ability to suppress PVCs on a Holter monitor, a surrogate endpoint, not on a trial measuring whether they saved lives. Physicians of my generation were taught in training that suppressing the marker was a reasonable surrogate for reducing the risk it marked, a conclusion that followed from a plausible mechanism rather than from outcome data, and the drugs were prescribed accordingly. This predates the guideline-development infrastructure the rest of this post describes. What it shows is that the structural problem, an unproven assumption reaching widespread practice through an official mechanism before being tested, doesn’t require a formal guideline to happen. A regulatory approval granted on the strength of a surrogate endpoint, reinforced by a plausible mechanism and ordinary clinical teaching, did the same job.

The Cardiac Arrhythmia Suppression Trial (CAST) tested that assumption directly in 1,498 post-MI patients with reduced ejection fraction and PVCs, randomized to encainide, flecainide, or placebo. The trial was stopped early. Arrhythmic or cardiac-arrest deaths occurred in 43 patients on active drug versus 16 on placebo (P = .0004); non-arrhythmic cardiac deaths occurred in 17 versus 5 (P = .01). [4] The explanation turned out to be proarrhythmia: these class I antiarrhythmic agents could themselves trigger new, often lethal, ventricular arrhythmias in a heart already damaged by infarction, an effect the pre-trial evidence had not been positioned to catch. Suppressing the marker on an ECG had been treated as equivalent to reducing the risk it marked, and in a damaged heart, the same drug class that could silence a PVC could also set off the rhythm that kills. CAST is credited with changing how cardiologists thought about antiarrhythmic drugs and about surrogate markers generally, not just about this one drug class.

In hindsight, the guideline movement that followed in the 1990s and 2000s would have done well to take CAST’s warning more seriously. A mechanism that makes sense is not evidence, and formalizing a plausible but untested assumption into a widely followed recommendation carries the same risk whether it takes the form of an FDA approval, a textbook teaching point, or, as the rest of this post documents, a published guideline.

A Guideline Clause Nobody Actually Vetted

Once clinical practice guidelines became medicine’s dominant vehicle for this kind of recommendation, the same failure didn’t need a new drug class and a halted trial to repeat itself. It could happen in the time it takes a committee to run out of time at a meeting. The clearest illustration of that mechanism I know of isn’t a published study. It’s an anecdote Brad Spellberg told on a podcast episode built around the essay above, and I include it here because it shows the mandate mechanism in its starkest, most literal form, not because it proves anything by itself. [2]

Spellberg, then an inpatient medicine director, was reviewing a core-measure fallout at his hospital: a patient with severe community-acquired pneumonia requiring ICU care, who was allergic to beta-lactam antibiotics, had been treated with vancomycin plus another agent. A state-level core-measure reviewer flagged the case as noncompliant, because the applicable pneumonia guideline specified that beta-lactam-allergic patients going to the ICU should receive aztreonam, a drug with no activity against the gram-positive organisms that cause most community-acquired pneumonia. Spellberg, an infectious disease physician, protested, then emailed three authors of the guideline to ask where the recommendation came from. The first said he didn’t remember it. The second said he would never use aztreonam for that indication. The third, the guideline’s first author, told him the committee had run out of time at the meeting before it could discuss that part of the table, so “it never got vetted.”

I want to be precise about what this anecdote does and doesn’t show. It is a single, secondhand account, recounted years after the fact on a podcast, of one guideline table and one reviewer’s compliance decision. I have not independently verified which guideline, committee, or year is being described, and I am not presenting it as evidence that guideline panels routinely skip vetting their own recommendations. What it illustrates, concretely, is the mechanism the rest of this post documents with harder data: a line in a table, however it got there, can become an enforceable standard, with a state reviewer and a hospital’s compliance score behind it, before anyone has checked it against evidence, or even reread it.

When Guidelines Get Tested Later

Most guideline recommendations are never subjected to the kind of large, definitive trial that would settle whether following them actually improves outcomes. Some eventually are. A 2013 analysis in Mayo Clinic Proceedings examined 1,344 original research articles published in the New England Journal of Medicine between 2001 and 2010. Of these, 363 tested an established medical practice rather than a new one. Of those 363, 146 (40.2 percent) found the existing practice to be no better than a lesser or no intervention, a reversal; 138 (38.0 percent) reaffirmed it; and the remainder were inconclusive. [3]

This figure describes practices that happened to get a rigorous retest during one ten-year window in one journal, not a random sample of every guideline recommendation in medicine, and it should not be read as “40 percent of guidelines are wrong.” What it documents is that among practices that were tested a second time with a real trial, wrong turns out to be common enough that the absence of routine retesting should worry us. The next four cases describe what that absence has looked like in practice, each one a recommendation that was adopted, enforced, and in most cases caused documented harm, before it was tested against the assumption it was built on.

Hormone Therapy for Heart Disease Prevention

This case, unlike CAST, belongs squarely to the guideline era, and the guideline itself is more self-aware than you might expect. In 1992, the American College of Physicians published a formal clinical guideline on counseling postmenopausal women about preventive hormone therapy, in the same Annals of Internal Medicine issue as the systematic evidence review it was based on. It recommended that women with coronary heart disease, or at increased risk for it, were “likely to benefit from hormone therapy,” citing “extensive and consistent evidence” of roughly a 35 percent reduction in coronary heart disease risk. [5,6] The guideline’s authors did not hide the nature of that evidence: “Estimates of the potential risks and benefits of hormone therapy come from results of observational studies. Randomized trials are required to prove these effects and to define their magnitude, but such trials would need to be very large and to continue for many years. The following recommendations are based on the best available current data but could change as the results of randomized trials become available.” [6] The trial they were describing, before it existed, was the Women’s Health Initiative, which randomized 16,608 postmenopausal women with a uterus to combined estrogen plus progestin or placebo to test that assumption directly.

The trial was stopped after a mean of 5.2 years of follow-up. Compared with placebo, hormone therapy was associated with a hazard ratio of 1.29 (95% CI, 1.02-1.63) for coronary heart disease, 1.41 (1.07-1.85) for stroke, 1.26 (1.00-1.59) for invasive breast cancer, and 2.13 (1.39-3.25) for pulmonary embolism. It also reduced the hazard ratio for colorectal cancer to 0.63 (0.43-0.92) and for hip fracture to 0.66 (0.45-0.98). [7] The investigators concluded that the overall harms outweighed the benefits for the purpose of primary disease prevention in this population, reversing more than a decade of guidance built on observational associations rather than a randomized test. The 1992 guideline had explicitly flagged that possibility and recommended hormone therapy anyway, in the absence of anything better to go on. That is precisely the mandate mechanism at work: a recommendation issued as provisional, pending a trial, nonetheless became the standard millions of women and their physicians acted on for the decade it took that trial to be run.

A Number Enforced for a Decade: The Vancomycin Trough Target

In 2009, a joint consensus guideline from the American Society of Health-System Pharmacists, the Infectious Diseases Society of America, and the Society of Infectious Diseases Pharmacists gave clinicians a specific, easy-to-apply target for dosing vancomycin in serious MRSA infections: a trough serum concentration of 15 to 20 mg/L. The trough was meant as a practical bedside stand-in for the actual pharmacodynamic driver of efficacy, a 24-hour area-under-the-curve-to-MIC ratio of at least 400, which was hard to calculate at the bedside in 2009. [8] That trough target was absorbed into hospital pharmacy dosing protocols and order sets widely, and into the standard of care clinicians were taught and expected to follow, well short of the kind of explicit regulatory or payer enforcement documented for the pneumonia and sepsis measures in the previous post, but a mandate in the medicolegal sense the Annals essay describes.

The 2009 guideline’s own authors later acknowledged that the 15-to-20 target had been “derived largely from retrospective, single-center, observational studies,” most of them studies of MRSA bloodstream infection specifically, not the broader range of serious infections to which hospitals then applied it. [9] The guideline itself remained formally unrevised for 11 years, even as a number of studies evaluating its real-world impact accumulated in the interim. [9] A 2020 revision, now also including the Pediatric Infectious Diseases Society, concluded plainly: “Trough-only monitoring, with a target of 15 to 20 mg/L, is no longer recommended based on efficacy and nephrotoxicity data in patients with serious infections due to MRSA.” [9] A retrospective study of 1,280 hospitalized patients cited in that revision found that dosing guided by estimated AUC, rather than trough alone, was associated with significantly less acute kidney injury (odds ratio, 0.52; 95% CI, 0.34-0.80; P = .003). Other data summarized in the same guideline linked trough and AUC values at the high end of the old target range to substantially higher rates of kidney injury, without a clearly demonstrated gain in efficacy to offset it. [9]

I want to apply the same caveat here that applies elsewhere in this post: the replacement standard, a daily AUC of 400 to 600, is itself still derived substantially from the kind of retrospective, single-center, observational data the 2020 guideline criticizes in the original trough target, a limitation its own authors state directly rather than concealing. [9] This is not a story of weak evidence giving way to strong evidence. It is a story of a specific, numeric target, built widely into pharmacy dosing protocols and treated as the standard of care, remaining the official recommendation for more than a decade while evidence against it accumulated, before it was formally withdrawn, and replaced in turn by a number that carries some of the same evidentiary limits its own authors openly acknowledge.

Treating Anemia in Kidney Disease: A Case Where the Correction Was Fast

Patients with chronic kidney disease become anemic, and observational studies through the 1990s and early 2000s consistently found that patients with higher hemoglobin levels had better cardiovascular outcomes than patients with lower ones. Erythropoiesis-stimulating agents (ESAs) could raise hemoglobin pharmacologically, and the reasonable-sounding extension of the observational data was that raising a CKD patient’s hemoglobin toward the normal range, not just out of the anemic range, should produce the same cardiovascular benefit seen in patients who were naturally less anemic.

Two trials published together in the same November 2006 issue of the New England Journal of Medicine tested that extension directly. CHOIR randomized 1,432 patients with non-dialysis CKD to a hemoglobin target of 13.5 g/dL or 11.3 g/dL using epoetin alfa. The composite outcome of death, myocardial infarction, hospitalization for heart failure, or stroke occurred in 17.5 percent of the higher-target group versus 13.5 percent of the lower-target group (hazard ratio, 1.34; 95% CI, 1.03-1.74; P = .03), and the trial was stopped early. [10] CREATE, randomizing CKD patients to a normal versus subnormal hemoglobin target with epoetin beta, found no cardiovascular benefit from full normalization and a higher rate of dialysis initiation in the normalized group. [11] A third trial, TREAT, published in 2009, randomized 4,038 patients with type 2 diabetes, CKD, and anemia to darbepoetin alfa targeting a hemoglobin near 13 g/dL or placebo with rescue therapy only if hemoglobin fell below 9. The composite of death or a cardiovascular event was not significantly different (31.4% vs. 29.7%; HR, 1.05; 95% CI, 0.94-1.17), but stroke occurred in 5.0 percent of the darbepoetin group versus 2.5 percent of the placebo group (HR, 1.92; 95% CI, 1.38-2.68; P < .001). [12]

This is the one case in this post where the correction moved fast. The FDA added a boxed warning narrowing the recommended hemoglobin target range within about six months of the CHOIR and CREATE publications, and clinicians in nephrology widely describe pulling back from higher hemoglobin targets at the bedside well before that formal label change took effect, on the strength of the trial data itself. I want to be precise about what is and is not documented here: a later analysis of national prescribing data found that the population-level decline in ESA dosing after the subsequent 2009 TREAT trial was gradual and uneven across insurance groups, with the clearest shift following a further FDA label revision in 2011, not an immediate drop after the trial’s publication. [13] Individual clinical practice and population-level prescribing data do not always move at the same speed, and I don’t have a trial-grade measurement of how fast bedside practice changed after CHOIR specifically. What is well documented is that, unlike RVU updates or most quality measures, a specific, clearly wrong target had an actual trial behind both its adoption and its reversal, and the reversal came in months rather than decades.

Episiotomy: A More Complicated Reversal

Not every reversal is a clean story of harm, and this one is worth including for that reason. Routine episiotomy at vaginal delivery was long practiced and recommended on the assumption that a controlled surgical cut prevented worse, uncontrolled tearing. A Cochrane review of 8 trials and 5,541 women comparing restrictive with routine episiotomy use found no difference in the rate of severe perineal or vaginal trauma overall (risk ratio 0.92; 95% CI, 0.72-1.18). Restrictive use was associated with less severe posterior perineal trauma (RR 0.67; 0.49-0.91) and less suturing and fewer healing complications, but more anterior perineal trauma (RR 1.84; 1.61-2.10) than routine use. [14]

The lesson here is not that episiotomy was proven harmful. It is that a practice justified for decades by an assumption about which injury it prevented was never tested against that specific assumption until relatively late, and when it was tested, the answer turned out to be genuinely mixed rather than a simple vindication or a simple reversal. The American College of Obstetricians and Gynecologists formally reversed its own guidance on this basis, issuing a 2006 practice bulletin recommending restrictive rather than routine use, citing this evidence. [15] Documenting a complicated result honestly is as much the point of this series as documenting a harmful one.

What This Selection Can and Cannot Show

These six cases, the aztreonam anecdote and five tested-and-reversed recommendations, were chosen because each illustrates how an unproven assumption reaches an official, enforceable status, whether through a guideline, a hospital compliance measure, or an FDA approval granted on a surrogate endpoint, and because, with the single noted exception, the underlying data are well documented. They are not a random sample of guideline recommendations. Most guideline recommendations, including good ones, are never subjected to a trial capable of overturning them, which means an unknown number of currently endorsed practices sit exactly where encainide, hormone therapy, and the vancomycin trough target sat before they were tested. This post cannot tell you which current recommendations are in that position. It can tell you that the process that would routinely check has not existed, and that when it does exist, as it did for anemia management in CKD, correction can happen quickly, and as it eventually did for the vancomycin target, correction can still take over a decade.

What Was Missing

Across the three posts in this series, the pattern is the same shape wearing different clothes. The first post described RVUs updated by a committee weighted toward the subspecialties whose income depended on the answer. The third described quality measures that reached national scale before anyone tested whether following them improved the outcome they claimed to improve, including the four-hour pneumonia rule and SEP-1, both of which began life as guideline recommendations. This post has documented the mechanism that connects guidelines to that third post’s quality measures: a guideline does not stay a recommendation once published, it becomes the medicolegal and regulatory standard against which performance is measured, with no requirement that it be tested first.

What this post has not yet examined is the guideline development process itself, the question of who sits on these panels, how conflicted they are, and what it means that two guideline committees looking at the same evidence on the same clinical question can reach opposite conclusions. That is the subject of the next post, and it is a related but distinct failure from the one documented here: this post asked whether a guideline’s recommendation was ever tested against an outcome; the next asks whether the process that produced the recommendation in the first place met even the field’s own standards for trustworthiness.

Bob’s Pearls

1. A guideline does not stay a suggestion once it is published. It becomes a medicolegal standard of care, a compliance target for regulators and payers, and a template for hospital order sets, whether or not its authors intended that.

2. A guideline table can become an enforceable standard before anyone has actually vetted it. In one secondhand account, a hospital was flagged as noncompliant by a core-measure reviewer for not using a drug with no activity against the organisms it was meant to treat, because a guideline committee ran out of time to discuss that part of the table before publishing it.

3. When established practices are put through a rigorous retest, a substantial share reverse. Of 363 NEJM articles testing an established practice between 2001 and 2010, 146 (40.2 percent) found the practice no better than a lesser or no intervention.

4. Suppressing a surrogate marker is not the same as treating the outcome it stands in for. The CAST trial found that suppressing PVCs after MI with encainide or flecainide increased both arrhythmic and non-arrhythmic cardiac deaths compared with placebo.

5. A guideline predicted its own overturning and recommended the practice anyway. The 1992 ACP hormone-therapy guideline said plainly that its recommendation rested on observational data and could change once a large randomized trial was run. It was right to worry: the Women’s Health Initiative, the trial it was describing before it existed, found the opposite of what a decade of guidance had assumed.

6. A specific numeric target was enforced in hospital pharmacy protocols for over a decade before it was reversed. The 2009 vancomycin trough target of 15 to 20 mg/L, based largely on retrospective single-center data, was formally withdrawn in 2020 for causing nephrotoxicity without a clear efficacy benefit, and its replacement carries some of the same evidentiary limitations.

7. This series’ one case of a fast correction. When CHOIR and CREATE showed that targeting a near-normal hemoglobin in CKD patients increased cardiovascular events and offered no benefit, the FDA narrowed its labeling within about six months, and nephrologists describe changing bedside practice even faster, well ahead of national prescribing data, which shifted more gradually.

8. Not every reversal is a clean story of harm. The episiotomy evidence shows a genuinely mixed result, less severe posterior trauma but more anterior trauma with restrictive use, which is a more honest outcome than either “vindicated” or “harmful” and deserves to be reported that way.

References

1. Spellberg B, Wright WF, Shaneyfelt T, Centor RM. The Future of Medical Guidelines: Standardizing Clinical Care With the Humility of Uncertainty. Ann Intern Med. 2021;174(12):1740-1742. doi:10.7326/M21-3034

2. Centor RM, Spellberg B, Shaneyfelt T. Annals On Call: Guidelines With Humility. Ann Intern Med. Podcast. doi:10.7326/A20-0026

3. Prasad V, Vandross A, Toomey C, et al. A decade of reversal: an analysis of 146 contradicted medical practices. Mayo Clin Proc. 2013;88(8):790-798. doi:10.1016/j.mayocp.2013.05.012

4. Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo. The Cardiac Arrhythmia Suppression Trial. N Engl J Med. 1991;324(12):781-788. doi:10.1056/NEJM199103213241201

5. Grady D, Rubin SM, Petitti DB, et al. Hormone therapy to prevent disease and prolong life in postmenopausal women. Ann Intern Med. 1992;117(12):1016-1037.

6. American College of Physicians. Guidelines for counseling postmenopausal women about preventive hormone therapy. Ann Intern Med. 1992;117(12):1038-1041. doi:10.7326/0003-4819-117-12-1038

7. Writing Group for the Women’s Health Initiative Investigators. Risks and benefits of estrogen plus progestin in healthy postmenopausal women: principal results from the Women’s Health Initiative randomized controlled trial. JAMA. 2002;288(3):321-333. doi:10.1001/jama.288.3.321

8. Rybak MJ, Lomaestro BM, Rotschafer JC, et al. Therapeutic monitoring of vancomycin in adult patients: a consensus review of the American Society of Health-System Pharmacists, the Infectious Diseases Society of America, and the Society of Infectious Diseases Pharmacists. Am J Health Syst Pharm. 2009;66(1):82-98. doi:10.2146/ajhp080434

9. Rybak MJ, Le J, Lodise TP, et al. Therapeutic monitoring of vancomycin for serious methicillin-resistant Staphylococcus aureus infections: a revised consensus guideline and review by the American Society of Health-System Pharmacists, the Infectious Diseases Society of America, the Pediatric Infectious Diseases Society, and the Society of Infectious Diseases Pharmacists. Am J Health Syst Pharm. 2020;77(11):835-864. doi:10.1093/ajhp/zxaa036

10. Singh AK, Szczech L, Tang KL, et al. Correction of anemia with epoetin alfa in chronic kidney disease. N Engl J Med. 2006;355(20):2085-2098. doi:10.1056/NEJMoa065485

11. Drüeke TB, Locatelli F, Clyne N, et al. Normalization of hemoglobin level in patients with chronic kidney disease and anemia. N Engl J Med. 2006;355(20):2071-2084. doi:10.1056/NEJMoa062276

12. Pfeffer MA, Burdmann EA, Chen CY, et al. A trial of darbepoetin alfa in type 2 diabetes and chronic kidney disease. N Engl J Med. 2009;361(21):2019-2032. doi:10.1056/NEJMoa0907845

13. Vu K, Zhou J, Everhart A, et al. Uptake of evidence by physicians: de-adoption of erythropoiesis-stimulating agents after the TREAT trial. BMC Nephrol. 2021;22:284. doi:10.1186/s12882-021-02491-y

14. Carroli G, Mignini L. Episiotomy for vaginal birth. Cochrane Database Syst Rev. 2009;(1):CD000081. doi:10.1002/14651858.CD000081.pub2

15. American College of Obstetricians and Gynecologists. ACOG Practice Bulletin No. 71: Episiotomy. Obstet Gynecol. 2006;107(4):957-962.

Bottom Line: A new drug or device cannot reach a patient until it passes a structured, adversarial testing process designed to catch harm before approval. A performance measure can be endorsed and written into a national payment or reporting program without any requirement that its effect on patient outcomes be tested first. Performance measures are interventions on how physicians practice, adopted on the assumption that measuring and rewarding a behavior will improve care, and, as the cases in this post show, that assumption often goes untested with anything like the rigor we require before approving a drug that makes the same claim. Quality reporting also carries a real, measurable cost: in one national survey, practices in four common specialties spent an average of 785 hours per physician per year on it. In the cases reviewed in this post, where investigators went back and tested whether a measure or program actually improved outcomes, the results were mostly null, mixed, or contested, and a few measures had documented harms. This is not an argument that quality measurement never works. It is an observation that, in these cases, measures became national requirements first and were evaluated afterward, and that revision came from outside criticism rather than from a scheduled review.

I served for several years in the 2010s on the American College of Physicians’ performance measures committee, and this asymmetry was a recurring frustration from inside that work. A medication or device faces years of controlled study before it reaches a patient, specifically because regulators assume an intervention can cause harm as well as benefit. A performance measure goes through its own review, endorsement bodies like the National Quality Forum assess reliability and feasibility, but nothing in that process requires evidence that the measure improves the outcome it claims to improve before it takes effect. Insurers and other purchasers write performance measures into contracts and reporting programs on the premise that specifying how physicians practice will improve outcomes or lower cost. That premise is testable, and several of the cases below suggest it does not hold uniformly. What none of these programs has built in is a mechanism requiring that the possibility of harm be checked before the measure takes effect, the way it would be for a drug. The four-hour pneumonia rule and SEP-1, discussed below, are two cases where the harm was documented only after the measure was already in wide use.

The first post in this series described who updates the RVU scale. The second described how the time around each encounter has grown in ways the 1988 measurements never captured. This post turns to one contributor to that time: the reporting of quality measures. It asks the series question again. If a quality measure is an intervention on physician behavior, has it been held to anything like the standard we would apply to a new drug?

What Quality Reporting Costs

In a 2016 Health Affairs study, Casalino and colleagues estimated that practices in four common specialties spend an average of 785 hours per physician per year, and more than $15.4 billion nationally, dealing with the reporting of quality measures. [1]

A 2021 study of Medicare’s Merit-based Incentive Payment System (MIPS) interviewed leaders of 30 physician practices about the 2019 program year. The mean cost of participation was $12,811 per physician, and the mean time was 201.7 hours per physician per year. Of those hours, 53.6 were physician time, 99.2 were medical assistant and nursing time, 28.6 were executive administrator time, and 20.3 were other staff time. Physician time accounted for 54 percent of the total cost. [2] Small primary care practices averaged $18,466 per physician and medium primary care practices $13,631, compared with $4,107 for large multispecialty practices. [2] These subgroup figures come from a small number of practices in each category and are descriptive only.

Both figures are interview or survey estimates, not audited time studies, and the MIPS study was qualitative with 30 practices. They are the best available documentation, and they describe a cost in hours that no physician is reimbursed for.

What Testing Found, Program by Program

I have organized the evidence by program rather than by argument, and I have tried to report each finding as its authors did. Where a finding is contested, I say so.

Paying for Measured Performance

The largest early US test was Medicare’s Premier Hospital Quality Incentive Demonstration. Jha and colleagues compared 252 participating hospitals with 3,363 hospitals doing public reporting alone, examining 30-day mortality among more than 6 million patients with acute myocardial infarction, heart failure, pneumonia, or coronary artery bypass grafting from 2003 to 2009. They found no evidence that the program decreased 30-day mortality. [3] The authors suggested that the size of the incentives and the choice of targets may matter, so the result may not generalize to a differently designed program. [3]

A similar program in England showed a different early picture. Advancing Quality, introduced in the northwest region in 2008, was associated with an 18-month reduction in mortality. [4] The long-term follow-up found that hospitals in the program kept improving on the measures, but by 42 months the reduced mortality among participating hospitals was no longer significant (difference -0.1 percentage points; 95% CI, -0.6 to 0.3). [5] A 2019 Cochrane review of hospital pay for performance concluded that the programs made little or no difference to overall mortality, with condition-specific results that were partly conflicting and inconsistent across studies. [6]

Door-to-Balloon Time

Guidelines recommend a door-to-balloon time of 90 minutes or less for patients with ST-elevation myocardial infarction undergoing primary PCI, and the interval became a performance measure. Menees and colleagues analyzed 96,738 admissions at 515 hospitals from July 2005 through June 2009. National door-to-balloon times improved, but in-hospital mortality was essentially unchanged, and after adjustment there was no significant association between the annual reduction in door-to-balloon time and mortality (odds ratio for a 10-minute reduction, 1.04). [7]

This is a national trend analysis, not a test of whether time to reperfusion matters for an individual patient. I am flagging that as my own reading of the design, not the authors’ conclusion. What the study documents is that improving the measured number did not produce a measurable change in national in-hospital mortality over that period.

The Four-Hour Pneumonia Rule

The 2003 Infectious Diseases Society of America guidelines for community-acquired pneumonia recommended starting antibiotics within 4 hours of hospitalization, and the recommendation became a quality indicator that some payers tied to hospital incentive compensation and that was publicly reported. [10]

Within a few years, published studies described the consequences. Kanwar and colleagues reported misdiagnosis of pneumonia and inappropriate antibiotic use as side effects of the 4-hour rule. [8] Welker and colleagues found that efforts to meet the measure reduced diagnostic accuracy. [9] In 2008, Wachter and colleagues reviewed the experience under the title “Lessons from a Flawed Performance Measure” and noted that the measure had recently been revised from 4 to 6 hours in response to criticism. [10]

SEP-1

The Severe Sepsis and Septic Shock Management Bundle (SEP-1) has required US hospitals to report adherence since October 2015. The 3-hour bundle requires a lactate measurement, blood cultures, broad-spectrum antibiotics, and 30 mL/kg or more of intravenous crystalloid for patients with hypotension or a lactate of 4.0 mmol/L or higher, all within 3 hours of “time zero,” a defined starting point for sepsis. [11]

The Infectious Diseases Society of America raised concerns about SEP-1 in a position paper, including sepsis overdiagnosis, conflation of sepsis and septic shock in the urgency of antibiotics, and pressure toward rapid judgments in stable patients who might turn out to have noninfectious syndromes. [12] In 2025, Ford and colleagues published a systematic review of 17 observational studies. Of 12 studies assessing the relationship between SEP-1 compliance and mortality, 5 showed a statistically significant benefit and 7 did not. None of the studies was at low risk of bias, and the authors found no moderate- or high-level evidence that compliance or implementation was associated with improved sepsis mortality. [13] One included interrupted time series across 11 hospitals (51,810 patients) found no difference in mortality trends before and after implementation (expected mortality difference 0.1%; 95% CI, -0.9% to 1.1%). [13]

Absence of high-quality evidence of benefit is not evidence that prompt sepsis care is unimportant. The review addressed the bundle as a compliance measure, not the value of early antibiotics for a patient in shock. In the FY2024 Hospital Inpatient Prospective Payment System final rule, released August 1, 2023, CMS finalized a plan to move SEP-1 into Hospital Value-Based Purchasing, a pay-for-performance program, beginning in fiscal year 2026. [24] Two months later, a joint position paper from IDSA, ACEP, and five other societies argued against that shift, citing multiple time-series analyses showing SEP-1 implementation increased broad-spectrum antibiotic use and lactate testing without improving the combined outcome of hospital death or discharge to hospice, and recommended retiring SEP-1 rather than converting it to a pay-for-performance measure. [25]

Readmission Penalties

The Hospital Readmissions Reduction Program (HRRP), established in 2010, began penalizing hospitals with higher-than-expected 30-day readmissions for heart failure, myocardial infarction, and pneumonia in 2012. [14]

Wadhera and colleagues studied Medicare fee-for-service beneficiaries aged 65 and older hospitalized for those conditions. For heart failure, 30-day postdischarge mortality was already rising before the program was announced (a 0.27% increase between the first two study periods). Announcement (a 0.49% increase) and implementation (a 0.52% increase) were each associated with a further increase relative to that baseline trend (difference in change, 0.22%, P = .01, and 0.25%, P = .001). [14] The absolute differences were small, on the order of a quarter of one percent. A separate analysis of a heart failure registry linked to Medicare claims found that implementation was temporally associated with reduced 30-day and 1-year readmissions and increased 30-day and 1-year mortality. [15]

These findings are contested. MedPAC’s 2019 update reported that raw heart failure mortality rose through 2013, while risk-adjusted mortality declined across all conditions from 2010 through 2017, with part of that decline reflecting coding changes. [16] A 2019 JACC: Heart Failure position paper stated that no sufficiently granular evidence exists to separate the program’s effects from other causes, while noting that heart failure mortality has risen since implementation. [17]

The question was already in print before penalties began. A VA analysis of 50,125 first heart failure hospitalizations from 2002 to 2006 found that 30-day mortality fell from 7.1% to 5.0% while 30-day heart failure rehospitalization rose from 5.6% to 6.1%, and its authors noted implications for using rehospitalization as a quality measure. [18] A July 2010 NEJM letter examined risk-adjusted readmission and death after heart failure hospitalization across 3,857 hospitals and asked whether readmission for any reason is the most appropriate index of poor quality. [19] This does not show that anyone predicted the specific finding above. It shows that the relationship between readmission and mortality was an open question when the penalties started.

MIPS in Primary Care

The most directly relevant test for primary care physicians is of MIPS itself. Bond and colleagues analyzed 80,246 primary care physicians participating in MIPS in 2019, with 3.4 million attributed patients. They compared MIPS scores with five process measures (diabetic eye examination, HbA1c screening, breast cancer screening, tobacco screening, and influenza immunization) and six adjusted outcome measures, including hospitalizations, emergency department visits, and ambulatory care-sensitive admissions. A process measure counts whether a recommended action was done; an outcome measure counts what happened to the patient. [20]

The authors concluded that MIPS scores were inconsistently related to performance on process and outcome measures, and that physicians caring for more medically complex and socially vulnerable patients were more likely to receive low scores even when they delivered relatively high-quality care. [20] Of the 80,246 physicians, 69,322 (86%) had high scores (above 75), 6,151 (8%) had medium scores, and 4,773 (6%) had low scores (30 or below). [20] The authors offered several possible explanations, including physicians’ freedom to choose which measures to report, measures of uncertain validity, and a program design that may reward the capacity to collect and report data. [20]

A separate analysis of first-year MIPS in 38,830 physicians found that quality scores were not associated with hospital rates of postoperative complications and were not generally associated with failure-to-rescue rates; low scores for cardiac surgeons were associated with higher hospital coronary artery bypass graft mortality and readmissions. The authors described the evidence for the validity of MIPS as limited. In that program, physicians could choose any 6 measures from 271 available. [21]

A third analysis addressed the measure list itself. Between 2017 and 2023, 137 of 275 MIPS quality measures (49%) were designated “topped out,” meaning reporting clinicians consistently achieve high performance, which triggers scoring caps and eventual removal. More than half of those topped-out measures were reported by fewer than 5 percent of eligible physicians at the time of designation. The authors concluded that topped-out designations may not reflect universally high performance. [22]

What This Selection Can and Cannot Show

I chose these cases because they are among the best documented, not because they are a random sample of measures. I did not find a verified count of how many measures across all payers have outcome evidence, and I am not claiming that most measures lack it. Some measures probably improve outcomes, and this post does not identify them.

Three further cautions apply. First, several of the null findings come from observational designs, and a null association can reflect co-interventions. A correspondent on the Jha study raised the point that private-payer incentives may have been operating in both the participating and comparison hospitals. [23] Second, mortality and hospitalization are blunt endpoints. A process measure can change practice without changing mortality, and a patient-experience measure was never aimed at mortality at all. Third, the England program shows that an effect can appear early and fade, which is a different problem from no effect.

What Was Missing

Across these cases, the recurring pattern is sequence. The 4-hour rule, the 90-minute door-to-balloon target, and the readmission penalties reached national scale before the evaluations above existed. Revision of the pneumonia measure followed published criticism. The proposed retirement of SEP-1 came from professional societies [25], in the same period that CMS moved to tie it to payment [24]. In none of these cases did I find a scheduled review that retired or revised a measure on the strength of outcome data.

In 2008, the authors of the pneumonia review recommended that quality measurement programs build in mechanisms to reassess measures over time, that representative end users participate in measure development, and that financial and intellectual biases in measure development be minimized. [10] Two of those recommendations echo the governance question from the first post in this series. The next post turns to guidelines, where several of the measures above began: the 4-hour pneumonia target and the 90-minute door-to-balloon target both started as guideline recommendations.

Bob’s Pearls

1. Quality reporting consumes measurable time. Practices in four common specialties averaged 785 hours per physician per year on measure reporting, and MIPS participation averaged 201.7 hours and $12,811 per physician in a 30-practice interview study.

2. Paying for measured performance has a thin outcome record. The largest US hospital program found no decrease in 30-day mortality, England’s early mortality reduction was no longer significant at 42 months, and a Cochrane review found little or no difference in overall mortality.

3. Improving the number is not the same as improving the outcome. National door-to-balloon times fell while in-hospital mortality stayed essentially unchanged, with no significant association between the two (odds ratio 1.04 per 10-minute reduction).

4. Time-based measures can distort diagnosis. The 4-hour pneumonia rule was associated with misdiagnosis and reduced diagnostic accuracy before it was revised to 6 hours. For SEP-1, 5 of 12 compliance studies showed benefit, none was at low risk of bias, and no moderate- or high-level evidence was found.

5. Readmission penalties remain contested. Two cohort analyses associated the program with higher heart failure mortality, while MedPAC’s risk-adjusted data showed a decline across conditions, and the question of cause remains unresolved.

6. MIPS scores do not track clinical performance consistently. Among 80,246 primary care physicians, 86 percent had high scores, and physicians with more medically complex and socially vulnerable patients were more likely to score low even when their care looked comparable.

References

1. Casalino LP, Gans D, Weber R, et al. US physician practices spend more than $15.4 billion annually to report quality measures. Health Aff (Millwood). 2016;35(3):401-406. doi:10.1377/hlthaff.2015.1258

2. Khullar D, Bond AM, O’Donnell EM, Qian Y, Gans DN, Casalino LP. Time and financial costs for physician practices to participate in the Medicare Merit-based Incentive Payment System: a qualitative study. JAMA Health Forum. 2021;2(5):e210527. doi:10.1001/jamahealthforum.2021.0527

3. Jha AK, Joynt KE, Orav EJ, Epstein AM. The long-term effect of Premier pay for performance on patient outcomes. N Engl J Med. 2012;366(17):1606-1615. doi:10.1056/NEJMsa1112351

4. Sutton M, Nikolova S, Boaden R, Lester H, McDonald R, Roland M. Reduced mortality with hospital pay for performance in England. N Engl J Med. 2012;367(19):1821-1828. doi:10.1056/NEJMsa1114951

5. Kristensen SR, Meacock R, Turner AJ, et al. Long-term effect of hospital pay for performance on mortality in England. N Engl J Med. 2014;371(6):540-548. doi:10.1056/NEJMoa1400962

6. Mathes T, et al. Pay for performance for hospitals. Cochrane Database Syst Rev. 2019. doi:10.1002/14651858.CD011156.pub2

7. Menees DS, Peterson ED, Wang Y, et al. Door-to-balloon time and mortality among patients undergoing primary PCI. N Engl J Med. 2013;369(10):901-909. doi:10.1056/NEJMoa1208200

8. Kanwar M, et al. Misdiagnosis of community-acquired pneumonia and inappropriate utilization of antibiotics: side effects of the 4-h antibiotic administration rule. Chest. 2007;131:1865-1869.

9. Welker JA, Huston M, McCue JD. Antibiotic timing and errors in diagnosing pneumonia. Arch Intern Med. 2008;168(4):351-356. doi:10.1001/archinternmed.2007.84

10. Wachter RM, Flanders SA, Fee C, Pronovost PJ. Public reporting of antibiotic timing in patients with pneumonia: lessons from a flawed performance measure. Ann Intern Med. 2008;149(1):29-32. doi:10.7326/0003-4819-149-1-200807010-00007

11. Rhee C, Yu T, Wang R, et al. Association between implementation of the Severe Sepsis and Septic Shock Early Management Bundle performance measure and outcomes in patients with suspected sepsis in US hospitals. JAMA Netw Open. 2021;4(12):e2138596. doi:10.1001/jamanetworkopen.2021.38596

12. Infectious Diseases Society of America position paper: recommended revisions to the National Severe Sepsis and Septic Shock Early Management Bundle (SEP-1) sepsis quality measure. Clin Infect Dis. 2021;72(4):541.

13. Ford JS, Morrison JC, Kyaw M, et al. The effect of Severe Sepsis and Septic Shock Management Bundle (SEP-1) compliance and implementation on mortality among patients with sepsis: a systematic review. Ann Intern Med. 2025;178(4):543-557. doi:10.7326/ANNALS-24-02426

14. Wadhera RK, Joynt Maddox KE, Wasfy JH, Haneuse S, Shen C, Yeh RW. Association of the Hospital Readmissions Reduction Program with mortality among Medicare beneficiaries hospitalized for heart failure, acute myocardial infarction, and pneumonia. JAMA. 2018;320(24):2542-2552. doi:10.1001/jama.2018.19232

15. Gupta A, Allen LA, Bhatt DL, et al. Association of the Hospital Readmissions Reduction Program implementation with readmission and mortality outcomes in heart failure. JAMA Cardiol. 2018;3(1):44-53. doi:10.1001/jamacardio.2017.4265

16. Medicare Payment Advisory Commission. Update: MedPAC’s evaluation of Medicare’s Hospital Readmission Reduction Program. December 2019. medpac.gov.

17. The Hospital Readmissions Reduction Program: nationwide perspectives and recommendations: a JACC: Heart Failure position paper. JACC Heart Fail. 2019. doi:10.1016/j.jchf.2019.07.012

18. Heidenreich PA, Sahay A, Kapoor JR, Pham MX, Massie B. Divergent trends in survival and readmission following a hospitalization for heart failure in the Veterans Affairs health care system 2002 to 2006. J Am Coll Cardiol. 2010;56(5):362-368. doi:10.1016/j.jacc.2010.02.053

19. Gorodeski EZ, Starling RC, Blackstone EH. Are all readmissions bad readmissions? N Engl J Med. 2010;363(3):297-298. doi:10.1056/NEJMc1001882

20. Bond AM, Schpero WL, Casalino LP, Zhang M, Khullar D. Association between individual primary care physician Merit-based Incentive Payment System score and measures of process and patient outcomes. JAMA. 2022;328(21):2136-2146. doi:10.1001/jama.2022.20619

21. Glance LG, et al. Association between the physician quality score in the Merit-Based Incentive Payment System and hospital performance in Hospital Compare in the first year of the program. JAMA Netw Open. 2021. pmc.ncbi.nlm.nih.gov/articles/PMC8335582

22. Chung YK, et al. Reporting rates of topped-out merit-based incentive payment system (MIPS) quality measures, 2017-2023. Health Aff Sch. 2026. doi:10.1093/haschl/qxag061

23. Correspondence: Premier pay for performance and patient outcomes. N Engl J Med. 2012. doi:10.1056/NEJMc1206170

24. Centers for Medicare & Medicaid Services. FY 2024 Hospital Inpatient Prospective Payment System Final Rule (CMS-1785-F). Federal Register. August 1, 2023.

25. Rhee C, Strich JR, Chiotos K, et al. Improving sepsis outcomes in the era of pay-for-performance and electronic quality measures: a joint IDSA/ACEP/PIDS/SHEA/SHM/SIDP position paper. Clin Infect Dis. 2024;78(3):505-513. doi:10.1093/cid/ciad447

Bottom Line: William Hsiao’s team measured the time and intensity of a bounded clinical encounter in 1988. What has grown enormously since then is largely unbounded, uncompensated time around the encounter, documentation, chart review, prior authorization, portal messages, and it has not grown evenly. It has concentrated in cognitive, longitudinal, medication-management practice, whether that practice is primary care or a cognitive subspecialty like infectious disease or nephrology, while largely sparing procedural subspecialties. This is a second, independent way the 1988 measurement has gone stale, separate from the RUC composition story in the previous post.

The previous post in this series argued that the RBRVS was built on rigorous 1988 methodology and that what has gone wrong since is an update process dominated by the subspecialties that benefit from its recommendations. This post picks up a different thread: even if the update process worked perfectly, the thing Hsiao’s team measured, physician time and intensity for a given service, has itself drifted since 1988 in ways the fee schedule has never caught up with.

What Hsiao Actually Measured

Hsiao’s surveys asked physicians to estimate the time, mental effort, technical skill, physical effort, and stress involved in performing a specific, definable service: an office visit, a procedure, a consultation. The unit of analysis was the encounter itself. This made sense in 1988, when a visit largely began when the patient arrived and ended when they left, and when the paper chart didn’t follow the physician home.

The Time Nobody Was Measuring

Electronic health records changed the shape of a physician’s day in ways no 1988 survey instrument could have anticipated. A 2020 study of roughly 100 million outpatient encounters across 155,000 physicians, using EHR activity logs rather than self-report, found physicians spending an average of just over 16 minutes actively using the EHR per encounter. [1] That average conceals wide variation by specialty. Primary care physicians averaged just under 20 minutes of active EHR time per encounter, among the highest of any specialty studied. [1, 2]

Here is the finding that should reframe how this series talks about the primary care problem: primary care was not alone at the top of that list, and the specialties keeping it company were not procedural ones. A national study of EHR metadata from over 200,000 ambulatory physicians using Epic systems, expressed as hours of EHR time per eight hours of scheduled patient time, found infectious disease at 8.4 hours, endocrinology at 7.7, nephrology at 7.5, and primary care (family and internal medicine) at 7.3. [3] At the opposite end: anesthesiology at 2.5 hours and orthopedics at 3.3. [3]

The fault line here is not primary care versus subspecialty medicine. It is cognitive, longitudinal, medication-management practice versus procedural practice. Infectious disease, endocrinology, and nephrology are subspecialties in every sense, board-certified, fellowship-trained, reimbursed at subspecialty rates for many services. They also carry documentation burdens comparable to or exceeding primary care’s. The study’s authors offered a mechanistic explanation consistent with this pattern: these specialties manage medically complex patients with multiple conditions, generating substantial visit-note documentation and high volumes of between-visit inbox communication. It is worth noting that this ranking comes from an Epic-based national dataset, a different EHR vendor and a different metric (hours per scheduled clinic time rather than minutes per encounter) than the Cerner-based study cited above, and the two converge on the same pattern: primary care and a cluster of cognitive subspecialties at the top, procedural specialties at the bottom.

Time-motion research in family medicine adds texture to the aggregate numbers. A study directly observing family physicians found they spent more time working in the EHR during clinic hours than they spent face-to-face with patients, and that most worked through lunch, stayed late, or took documentation home. [4] This is not physicians being slow. It is the structure of longitudinal, multi-problem visits generating more discrete data points to capture than a bounded procedural encounter does.

An Honest Gap: Prior Authorization

It would be tempting to extend this same pattern to prior authorization burden, and I want to resist that temptation rather than assert something the data doesn’t show. The American Medical Association’s most recent prior authorization survey found practices completing roughly 40 requests per physician per week and spending about 13 hours of physician and staff time on them, drawn from a sample that was 40 percent primary care and 60 percent subspecialists. [5] The published results do not break this burden out by specialty type. It is plausible that prior authorization, like documentation time, concentrates unevenly, medication-heavy subspecialties like rheumatology and oncology are frequently cited anecdotally as facing extremely high prior authorization volume for biologics and specialty drugs, which would cut against a simple primary-care-versus-everyone-else framing just as the EHR data does. But I don’t have the specialty-level breakdown to state this with confidence, and naming that gap honestly is more useful than filling it with a plausible-sounding guess.

Why This Belongs in the Same Series as the RUC Story

A second symptom of the same root problem, not a wholly separate one. The previous post described a governance failure: the body updating RVU values is structurally weighted toward the subspecialties whose income depends on those values. This post describes what that same absence of re-measurement looks like from a different angle: the thing being valued, physician time and intensity for a given service, has changed shape since 1988, and if the RUC’s surveys were re-run today with current data, they would presumably have to capture it. Both problems trace to the same missing piece: there is no equivalent of post-marketing surveillance checking whether the original measurement still holds, independent of who benefits from the answer. A drug’s dosing doesn’t get revisited only when the company that sells it requests a review. Physician work value should not depend entirely on a committee dominated by the specialties being valued, and it should not still be measuring 1988’s version of what a clinical encounter requires.

The next post in this series turns to quality measures, many of which are a direct contributor to the documentation burden described here, and asks whether the measures themselves have been tested for the outcomes they claim to improve.

Bob’s Pearls

1. Primary care physicians average close to 20 minutes of active EHR time per encounter, among the highest of any specialty measured, using EHR activity-log data rather than self-report.

2. The specialties matching or exceeding primary care’s documentation burden were cognitive subspecialties, infectious disease, endocrinology, and nephrology, not procedural ones. Anesthesiology and orthopedics showed documentation burdens roughly one-third as large.

3. The real fault line is cognitive, longitudinal practice versus procedural practice, not primary care versus subspecialty medicine as a category. This is a sharper claim than the usual framing and better supported by the data.

4. Prior authorization burden by specialty type remains an honest gap in the available data. Aggregate survey figures exist; a specialty-level breakdown does not, at least not from a source I can cite with confidence.

5. The 1988 Hsiao survey measured the time and intensity of a bounded encounter. Uncompensated time has grown substantially since then, concentrated unevenly, and no systematic mechanism has recaptured that drift in the work-value formula.

References

1. Overhage JM, McCallie D Jr. Physician Time Spent Using the Electronic Health Record During Outpatient Encounters: A Descriptive Study. Ann Intern Med. 2020;172(3):169-174. doi:10.7326/M18-3684

2. Healio Primary Care. Physicians spend 16 minutes using EHR for each patient encounter. Reporting on Overhage/McCallie, 2020.

3. Holmgren AJ, Sinsky CA, Rotenstein L, Apathy NC. National Comparison of Ambulatory Physician Electronic Health Record Use Across Specialties. J Gen Intern Med. 2024. doi:10.1007/s11606-024-08930-4

4. Young RA, et al. A Time-Motion Study of Primary Care Physicians’ Work in the Electronic Health Record Era. Fam Med. 2018;50(2):91-99.

5. American Medical Association. 2025 AMA Prior Authorization Physician Survey. ama-assn.org.

Bottom Line: In 1988, a Harvard research team led by William Hsiao did exactly what medicine should do before adopting a major intervention. They measured physician work systematically, tested their methodology, and built a fee schedule grounded in data rather than habit. The Resource-Based Relative Value Scale (RBRVS) that resulted from that work was sound science. What has gone wrong with physician payment since 1992 is not a failure of the original research. It is the absence of any mechanism, comparable to post-marketing surveillance for a drug, to check whether the system built on that research is still serving its intended purpose thirty years later.

This is the first post in a series asking a single question: why does medicine adopt major system-level interventions, in payment, in quality measurement, in guidelines, without the kind of ongoing testing we require before a new drug reaches a pharmacy shelf? RVUs are where the series starts, because the RVU story is more complicated, and more instructive, than the usual complaint that “subspecialists get paid more than primary care.” (Internal medicine and family medicine are specialties in their own right; the imbalance this post describes runs between primary care and procedural subspecialty medicine.) The problem was never really about how RVUs were created. It is about who was left in charge of updating them.

Before RBRVS: Charges Based on Charges

Before 1992, Medicare paid physicians based on “usual, customary, and reasonable” charges. In practice, this meant fees were largely set by what physicians had charged in the past, which meant they tracked whatever the market had settled on rather than any measure of the actual work involved. Procedural subspecialties, whose historical charges were high, stayed high. Cognitive work, evaluation, counseling, coordinating care, was chronically undervalued because it had always been undervalued. The system had no mechanism to correct itself.

Congress asked Harvard economist William Hsiao and his team to fix this. The resulting Resource-Based Relative Value Scale study, published across several 1988 papers, is a genuinely rigorous piece of health services research. [1] Hsiao’s group surveyed physicians directly about the time, mental effort, technical skill, physical effort, and stress involved in performing specific services, and used those surveys to build relative work values that could be compared across specialties. It was, by the standards of 1988 health policy research, a real measurement study, not an opinion poll and not a negotiation.

This is worth stating plainly: the creation of the RVU system was not a “sexy” idea adopted on faith. It was closer to the kind of rigorous baseline measurement we would want before intervening in any complex system. Congress adopted the RBRVS in the Omnibus Budget Reconciliation Act of 1989, with the fee schedule phased in starting January 1, 1992.

Then: Who Gets to Update It?

Here is where the story changes character. A relative value scale is not static. New procedures appear, existing ones change, and the values need periodic revision. The AMA created a body to handle this: the Relative Value Scale Update Committee, universally known as the RUC (pronounced “ruck”).

The RUC is a volunteer committee of roughly 31 physicians, nominated by specialty and subspecialty societies, that reviews services and recommends work values to CMS. [2] CMS is not obligated to accept these recommendations. In practice, it does, most of the time.

The composition of the RUC matters more than any other fact in this story. One analysis by the advocacy group Public Citizen found that primary care physicians made up as little as 7 percent of RUC membership, with the remainder weighted toward procedural subspecialty medicine. [3] This is the structural feature that should have triggered concern from day one: a committee dominated by the subspecialties whose members stand to benefit from the values it recommends, reviewing those same values, with a track record of high deference from the agency that is supposed to provide independent oversight.

The Acceptance-Rate Data, Stated Honestly

The figure that circulates most often in this discussion is that CMS accepts RUC recommendations “about 90 percent of the time.” The more precise number, from the most frequently cited analysis, is 87.4 percent, based on Miriam Laugesen and colleagues’ review of 2,768 work-value decisions between 1994 and 2010. [4] A separate Public Citizen analysis of a narrower 2012-2014 window found the acceptance rate had fallen to roughly 65 percent for that period, and, notably, that when CMS did deviate from a RUC recommendation, it lowered the value more than 90 percent of the time. [3] That second finding matters: it suggests the RUC’s recommended values skew high on average, not simply that CMS and the RUC disagree unpredictably.

I want to flag an honest complication here, because naming it is more useful than smoothing over it. The Laugesen analysis also found that when CMS did deviate from RUC recommendations, it was more likely to reduce values for surgical, radiologic, and other procedural subspecialty services relative to evaluation and management (E/M) services, the billing codes that cover most primary care visits. [4] Read narrowly, that finding cuts against the simplest version of the “RUC systematically shortchanges primary care through CMS’s deference” story. The authors themselves noted that this pattern does not explain why the income gap between primary care and procedural subspecialty medicine has persisted. In other words: the mechanism of harm to primary care is not simply “CMS rubber-stamps whatever the RUC says.” It is more accurately: the RUC process as a whole, from which services get reviewed at all to how work value surveys are constructed, has structurally favored procedural work over cognitive work for decades, even when CMS’s individual up-or-down votes on specific values show some corrective tendency at the margins.

The Numbers

The structural argument above is about mechanism. Here is the outcome, reported without a claim about what caused it.

Medscape’s annual physician compensation survey, based on more than 7,000 respondents across 29-plus specialties, reported average primary care compensation of $265,000 in 2022, $277,000 in 2023, and $281,000 in 2024. [5] Average subspecialty compensation over the same three years was $382,000, $394,000, and $398,000. [5] The dollar gap has held at $117,000 across all three survey years. Going back further, the same survey found primary care compensation rose from $195,000 in 2015 to $237,000 in 2022, a 21.5 percent increase, while subspecialty compensation rose from roughly $368,000 in 2021 to $398,000 in 2024, a smaller percentage increase on a much larger base. [5]

Separately, the American Medical Association, drawing on Federal Register data, Medicare Trustees’ Reports, Bureau of Labor Statistics figures, and Congressional Budget Office estimates, has tracked Medicare physician payment adjusted for inflation in practice costs. By that measure, Medicare physician payment declined 20 percent from 2001 to 2021, 26 percent by 2023, 29 percent by 2024, and 33 percent by 2025. [6] This erosion applies to the conversion factor that multiplies every RVU, primary care and procedural subspecialty codes alike, and reflects a different mechanism than RUC composition: the statutory absence of an inflation adjustment for the physician fee schedule.

These are observational figures, not a controlled comparison, and they do not by themselves establish that RUC composition or CMS deference caused the compensation gap or the payment erosion. Readers can weigh how much of each pattern the structural story above plausibly explains.

Budget Neutrality: Why This Is Zero-Sum

One more structural feature explains why RVU reform generates so much subspecialty-society resistance whenever primary care advocates ask for E/M codes to be revalued upward. Medicare physician payment operates under a budget neutrality requirement: increasing the total value of one set of services requires offsetting reductions somewhere else in the fee schedule. There is no mechanism to simply add money to fix an undervalued category. Every gain for primary care work has to come out of some other subspecialty’s payment, which is precisely why E/M revaluation efforts have historically triggered organized opposition rather than broad consensus.

What This Series Is Actually About

The RBRVS was built with real methodology. What has been missing since 1992 is any equivalent of post-marketing surveillance: an independent, ongoing check on whether the system built from that research is still functioning as intended, conducted by people who do not have a direct financial stake in the answer. A body that reviews and revalues services should not be composed overwhelmingly of the subspecialties whose income depends on those values. That is not a subtle design flaw. It was predictable from the moment the RUC was formed, and it remained unaddressed for over three decades.

The next posts in this series turn to two other places where medicine has adopted large-scale interventions, quality measures and clinical guidelines, without the ongoing evaluation we would demand of a new drug. The series closes with a concrete proposal for how that evaluation could actually work.

Bob’s Pearls

1. The RBRVS was built on real survey-based measurement by Hsiao’s team, not intuition. This series is not an argument that all systemic reform is reckless. The original 1988 work was methodologically sound.

2. CMS has agreed with RUC work-value recommendations 87.4 percent of the time by the most-cited analysis (1994-2010), a rate that fell closer to 65 percent in a later 2012-2014 window. When CMS disagrees, it lowers the value more than 90 percent of the time, suggesting the RUC’s recommendations skew high.

3. Primary care representation on the RUC has been estimated as low as 7 percent of its membership, a structural imbalance that was foreseeable from the committee’s creation.

4. Budget neutrality makes RVU reform zero-sum. Raising the value of E/M codes requires cutting value elsewhere in the fee schedule, which is the mechanical reason primary care payment reform generates subspecialty-society resistance rather than consensus.

5. The honest complication: CMS’s individual work-value decisions have shown some tendency to favor E/M over procedural codes at the margins, which means the harm to primary care is better explained by the RUC process as a whole than by a simple story of CMS rubber-stamping subspecialty-favoring recommendations.

6. The observational data: primary care and subspecialty compensation have held a stable $117,000 gap across 2022-2024, and Medicare physician payment has declined 33 percent since 2001 when adjusted for practice-cost inflation. These figures document the outcome; they don’t by themselves prove the structural story above is the cause.

References

1. Hsiao WC, Braun P, Dunn D, Becker ER, DeNicola M, Ketcham TR. Results and policy implications of the resource-based relative-value study. N Engl J Med. 1988;319(13):881-888. doi:10.1056/NEJM198809293191330

2. American Medical Association. An Introduction to the RUC. ama-assn.org.

3. Public Citizen. Overhaul Needed of System for Establishing Medicare Physician Payments. citizen.org.

4. Laugesen MJ, Wada R, Chen EM. In setting doctors’ Medicare fees, CMS almost always accepts the relative value update panel’s advice on work values. Health Aff (Millwood). 2012;31(5):965-972. doi:10.1377/hlthaff.2011.0557

5. Medscape Physician Compensation Report, 2023-2025 editions. medscape.com.

6. American Medical Association. Medicare Updates Compared to Inflation (2001-2025). ama-assn.org.

This conversation with Dr. David Felson, rheumatologist and epidemiologist at Boston University, already produced one post on this Substack: the piece on the cost-effectiveness of GLP-1 receptor agonists for knee osteoarthritis. That post covered the economics and the core pathophysiology reframing. What follows here is the rest of that same conversation, the part that didn’t fit there because it belongs with a different thread running through this series.

“It’s not a degenerative disease. It’s an inflammatory disease.” That is Dr. Felson, answering flatly when I summarized his own explanation back to him. This is the same claim this series has already made with Dr. Tuhina Neogi, from the same institution and the same osteoarthritis research community. That is worth naming plainly: this is not two isolated observers arriving at the same conclusion from separate premises, it is two colleagues in the same subfield reaching for the same reframing, which is a real signal that the field has moved but a weaker one than true independent confirmation would be.

The Colchicine Puzzle, From a Second Angle

In the Neogi post, low-dose colchicine and canakinumab, an IL-1β inhibitor, both reduced the rate of knee and hip replacement in cardiovascular trials never designed to study osteoarthritis. I asked Dr. Felson the obvious follow-up: what is colchicine actually doing to the joint? His answer was refreshingly candid. “We’re confused a little bit as to how it might be working.” Smaller, shorter trials testing colchicine directly in osteoarthritis patients have not shown much benefit, which sits awkwardly next to the large, long cardiovascular trial that did.

His best guess is the same one this series has been circling since the CPPD post: a substantial number of osteoarthritis patients may harbor microscopic calcium pyrophosphate crystals that are quietly triggering an inflammatory cascade, and colchicine, by inhibiting the NLRP3 inflammasome, may be blocking that cascade at its source rather than acting on osteoarthritis broadly. This is speculation, not established mechanism. Neither Dr. Felson nor Dr. Neogi presented it as proven, and no trial has yet tested colchicine specifically in a CPPD-confirmed osteoarthritis subgroup. Both rheumatologists work in the same osteoarthritis research community, so two of them raising the same hypothesis is worth reporting, not treated as independent confirmation. It still gives this series’ central thread, that a specific, crystal-driven, more inflammatory phenotype may explain results that look confusing when osteoarthritis is treated as one disease, a second voice.

Two Open Questions, Not One Mechanism

Worth setting next to the CPPD hypothesis rather than folded into it. The GLP-1 cost-effectiveness post already covered Dr. Felson’s point that fat mass itself, not just the load it adds, drives knee OA inflammation: patients with obesity had more hand pain than their hand OA severity explained, and bariatric surgery improved that hand pain even though hands bear no weight, an effect he attributed to falling leptin levels rather than reduced load. This is a separate, unconfirmed mechanism from the CPPD hypothesis above, not the same pathway restated. Crystal-triggered inflammasome activation and leptin-driven adipokine signaling are distinct biological processes. Colchicine and canakinumab may be interrupting the former; bariatric surgery, and presumably GLP-1-driven fat loss, may be reducing the latter. Both are plausible contributors to the more inflammatory end of the osteoarthritis spectrum, and nothing in either conversation ties them together into a single confirmed pathway.

A Caution Worth Attaching to the Prescription You’re Already Writing

One point from this conversation has not appeared elsewhere in this series and deserves its own mention. Weight loss on a GLP-1 receptor agonist is not pure fat loss. Some fraction of the weight lost is lean body mass, muscle, not just adipose tissue, and the muscles crossing the knee are precisely what distribute load and protect the joint when a patient carries excess weight. Losing weight typically reduces the load-bearing demand enough that this trade-off nets out favorably, but Dr. Felson flagged it as a genuinely open question rather than a settled reassurance: nobody has yet followed a large cohort of GLP-1 users with knee osteoarthritis to see whether meaningful lean mass loss offsets the benefit of reduced load in a subset of patients. This is not a reason to withhold GLP-1 therapy from an eligible patient. It is a reason to keep encouraging resistance training alongside it, and not to assume that weight loss alone fully substitutes for muscle strength around the joint.

Bob’s Pearls

1. A second rheumatologist in the same subfield makes the same correction: osteoarthritis is inflammatory, just milder than rheumatoid arthritis or lupus, not a passive degenerative process. Worth noting that this reflects a shift within one research community rather than two unrelated observers independently confirming it, a real signal of where the field is heading, but a more modest one than true independent confirmation.

2. Occult CPPD remains the leading, unproven hypothesis for why colchicine reduces joint replacement in osteoarthritis. Two rheumatologists in the same research community have now floated this, not independently. It has not been tested directly and should be described to learners as a hypothesis, not a mechanism.

3. A separate, equally unconfirmed hypothesis: fat mass itself, via leptin and other adipokines, may drive an inflammatory contribution to osteoarthritis distinct from crystal deposition. These are two different candidate mechanisms, not one pathway, and neither is established.

4. GLP-1-induced weight loss includes some lean mass loss, and muscle strength is what protects a loaded joint. Pair GLP-1 therapy with resistance training rather than treating weight loss alone as sufficient joint protection, an unresolved question that deserves longer-term study.

This podcast episode featured Dr. David Felson, rheumatologist and epidemiologist at Boston University’s Chobanian & Avedisian School of Medicine. The episode discussed is: Centor RM, Felson DT. “Annals On Call - Cost-Effectiveness of Weight Loss Drugs for Knee Osteoarthritis.” Ann Intern Med. 2025;178:e2505155OC. doi:10.7326/ANNALS-25-05155-OC [1]. This post connects directly to the earlier posts in this series on CPPD and on colchicine, canakinumab, and joint replacement with Dr. Tuhina Neogi, and does not repeat the cost-effectiveness analysis already covered in the separate GLP-1 series post built from this same episode.

References

[1] Centor RM, Felson DT. Annals On Call - Cost-Effectiveness of Weight Loss Drugs for Knee Osteoarthritis. Ann Intern Med. 2025;178:e2505155OC. Epub 2 December 2025. doi:10.7326/ANNALS-25-05155-OC

Osteoarthritis is not the inflammation-free disease many of us were trained to picture. It is a less inflammatory disease, not a non-inflammatory one, and that single word substitution changes what treatments are worth considering. Neither of the trials in this post was designed to study osteoarthritis. Both were cardiovascular trials, run in patients with heart disease, looking for heart outcomes. And both, almost by accident, found that an anti-inflammatory drug reduced the rate of knee and hip replacement. If a drug developed for a completely different organ system slows the disease that eventually sends a knee or a hip to the operating room, the “wear and tear” model of osteoarthritis cannot be the whole story.

My guest was Dr. Tuhina Neogi, Professor of Medicine and Chief of Rheumatology at Boston University’s Chobanian & Avedisian School of Medicine, whose research focuses on osteoarthritis and gout epidemiology and pain mechanisms [3]. I wanted her to walk through why two cardiology trials just handed rheumatology one of its more interesting open questions in years.

Two Trials That Weren’t Looking for This

The first is LoDoCo2, a cardiovascular trial of 5,522 patients with chronic coronary disease randomized to colchicine 0.5 mg daily or placebo. In an exploratory analysis, total knee or hip replacement occurred in 2.5% of the colchicine group versus 3.5% of the placebo group over a median follow-up of about 28.6 months, an incidence rate of 0.90 versus 1.4 per 100 person-years, translating to roughly a 30% relative risk reduction [1]. The absolute difference is one percentage point over roughly two and a half years, an NNT of about 100 to prevent one joint replacement over that period. Modest at the individual level, but osteoarthritis is the most common form of arthritis worldwide, affecting more than half a billion people, so a small per-patient effect scales into a meaningful public health signal.

The second is CANTOS, a cardiovascular trial of just over 10,000 patients randomized to canakinumab, an IL-1β monoclonal antibody, or placebo. A post hoc analysis found a similarly patterned reduction in knee or hip replacement across all three canakinumab doses, with a hazard ratio around 0.60, roughly a 40% relative risk reduction [2].

Colchicine and canakinumab do not share a drug class, but they converge on the same target. Colchicine inhibits the NLRP3 inflammasome, which blocks downstream production of IL-1β. Canakinumab blocks IL-1β directly. Two structurally unrelated drugs, tested in two unrelated cardiovascular populations, both suppressing the same cytokine, both showing the same signal in an outcome neither trial was built to measure. That convergence is what makes this worth taking seriously rather than dismissing as noise.

Degenerative Was the Wrong Word

I asked Dr. Neogi directly: how does a disease I was trained to call degenerative, and therefore assumed to be non-inflammatory, respond to anti-inflammatory drugs? Her answer reframed the question. “The term non-inflammatory is a misnomer. It’s just less inflammatory.” “Degenerative” describes wear accumulating passively over time; it does not describe an active process a drug could interrupt. Primary care physicians see the evidence against that model constantly and may not name it: a swollen knee with an effusion is a sign of inflammation, and imaging, ultrasound or MRI, will almost universally show some degree of synovitis or effusion in osteoarthritis. The degree simply falls well short of rheumatoid or psoriatic arthritis. Osteoarthritis, in her framing, is a disease of the whole joint, a reflection of repair mechanisms failing to keep pace with damage mechanisms, and cytokines sit inside that failure, not outside it.

Why Joint Replacement Is an Attractive, Imperfect Outcome

Osteoarthritis drug trials have a long history of failing to show benefit, and one open question is whether that reflects the wrong outcome measures and trial durations for a slow, progressive disease. Knee replacement is an appealing proxy for end-stage disease because roughly 97% of knee replacements are performed for knee osteoarthritis. Hip replacement is murkier, since hip fractures also drive replacement, though OA still accounts for the majority. Osteoarthritis is now the third leading hospital discharge diagnosis in the United States because of the sheer volume of elective joint replacements.

The outcome is also genuinely imperfect, and that imperfection matters for how these results should be read. The decision to replace a joint is driven substantially by symptoms, and it is also shaped by patient preference, surgeon preference, comorbidities, and insurance coverage. Neither trial can distinguish whether colchicine or canakinumab reduced replacement because symptoms improved, because joint structure was preserved, or both. That distinction is not academic: a drug cannot be labeled disease-modifying for osteoarthritis, from a regulatory standpoint, without demonstrated structural benefit. From a patient’s standpoint, symptomatic improvement alone would still matter enormously, since options for managing osteoarthritis pain remain limited.

Where This Connects Back to Crystal Arthritis

I asked Dr. Neogi the question this series has been circling: how much of what looks like plain osteoarthritis is actually undiagnosed crystal arthritis riding along underneath it, gout or CPPD contributing inflammation nobody is measuring. LoDoCo2 specifically excluded patients with known gout at enrollment and still found the same signal, which argues against occult gout as the explanation, though it cannot fully exclude undiagnosed cases.

Calcium crystal deposition is a bigger and more open question. By Dr. Neogi’s account, calcium crystal deposits are nearly universal at end-stage osteoarthritis, and whether those crystals actively drive symptoms or structural deterioration, rather than being an incidental finding of an already-damaged joint, remains genuinely unresolved. Her group has been using dual-energy CT, the same underlying technology used to identify urate deposits in gout but tuned to different energy levels, to identify calcium mineral deposition in osteoarthritic knees, with newer methods now attempting to distinguish CPPD from basic calcium phosphate crystals, though she was clear this is not yet a clinical tool. A paper from her group, accepted for publication at the time of this recording but not independently verified here since it has no citable journal or date yet, reportedly found that mineralization on CT was associated with increased risk of pain and pain worsening. This raises the possibility that the LoDoCo2 and CANTOS signals are picking up a specific, more inflammatory osteoarthritis phenotype tied to intra-articular mineralization, rather than an effect present across all osteoarthritis uniformly.

That possibility lines up directly with the pattern described in the CPPD post in this series: chronic, low-grade CPPD may present as an atypical distribution of osteoarthritis rather than an acute flare, and calcium crystal deposition and osteoarthritis are not two separate diagnoses so much as overlapping processes that current clinical tools mostly cannot separate.

Osteoarthritis Is Probably Several Diseases

Dr. Neogi’s broader point, and one she was careful to attribute to the osteoarthritis research community specifically rather than rheumatology as a whole, is that osteoarthritis likely represents multiple distinct pathways converging on a common radiographic and clinical endpoint. By the time x-ray changes are visible, the process producing them has typically been under way for years. A trial that lumps every osteoarthritis patient together and tests one mechanism, one molecule, is likely to dilute a real effect present in only a subset of patients, which may explain why so many prior osteoarthritis trials have been negative. LoDoCo2 and CANTOS found their signal only because of sheer size and long follow-up, not because either trial identified the right subgroup in advance. The next logical step, in her view, is a prospective trial with baseline imaging or biomarker data that can identify which phenotype, mineralization, effusion, synovitis, actually predicts response.

What This Means in Clinic Today

Dr. Neogi was direct that this is not a reason to start every osteoarthritis patient on colchicine. The first priority remains evidence-based basics that are inconsistently delivered: addressing weight in patients who are overweight or obese, since obesity is among the strongest risk factors for osteoarthritis and its progression, encouraging physical activity, and referring to physical therapy. Guideline-directed symptom management comes next.

Where she would personally consider colchicine, off-label, is a patient who has failed standard management and has chondrocalcinosis on radiograph, or recurrent effusions. She noted that LoDoCo2 used 0.5 mg once daily, a lower dose than the 0.5 mg twice daily used in some older, smaller colchicine trials, and that there is no reason to reach for the higher dose given the added side-effect burden. Colchicine 0.5 mg (branded as Lodoco) received FDA approval in June 2023, but specifically for reducing cardiovascular event risk in patients with established atherosclerotic disease or multiple risk factors; it does not carry an osteoarthritis indication, and this use remains off-label and investigational.

Bob’s Pearls

1. “Non-inflammatory osteoarthritis” is a misnomer. Synovitis and effusion are common on exam and near-universal on imaging in osteoarthritis; the inflammation is real, just milder in degree than rheumatoid or psoriatic arthritis.

2. Two cardiovascular trials, neither designed to study osteoarthritis, both found that suppressing IL-1β reduced knee and hip replacement, colchicine through NLRP3 inflammasome inhibition, canakinumab through direct IL-1β blockade. Two different drugs converging on the same signal in an outcome nobody was looking for is a stronger form of evidence than either trial alone.

3. Joint replacement is an attractive proxy for end-stage osteoarthritis, since about 97% of knee replacements are performed for it, but it cannot distinguish a symptomatic benefit from a structural one. That distinction determines whether a drug could ever be labeled disease-modifying.

4. Calcium crystal deposition is nearly universal at end-stage osteoarthritis, and whether it drives disease or just marks it is still unresolved. This may define a more inflammatory osteoarthritis phenotype, connecting directly to the CPPD pattern described earlier in this series.

5. Osteoarthritis is very likely several diseases sharing one endpoint, not one disease with variable severity. Trials that treat it as monolithic risk diluting real effects that exist only in a subgroup.

6. A reasonable off-label consideration, not a blanket recommendation: low-dose colchicine (0.5 mg once daily) in a patient who has failed standard osteoarthritis management and has chondrocalcinosis on x-ray or recurrent effusions. The FDA-approved indication for this dose is cardiovascular risk reduction, not osteoarthritis.

This podcast episode featured Dr. Tuhina Neogi, Professor of Medicine and Chief of Rheumatology at Boston University’s Chobanian & Avedisian School of Medicine. The articles discussed are: Heijman MWJ, Fiolet ATL, Mosterd A, et al. “Association of Low-Dose Colchicine With Incidence of Knee and Hip Replacements: Exploratory Analyses From a Randomized, Controlled, Double-Blind Trial.” Ann Intern Med. 2023;176. doi:10.7326/M23-0289 [1]. Schieker M, Conaghan PG, Mindeholm L, et al. “Effects of Interleukin-1β Inhibition on Incident Hip and Knee Replacement: Exploratory Analyses From a Randomized, Double-Blind, Placebo-Controlled Trial.” Ann Intern Med. 2020;173:509-515. doi:10.7326/M20-0527 [2]. Centor RM, Neogi T. “Annals On Call - Low-Dose Colchicine and Incident Knee and Hip Replacements.” Ann Intern Med. 2023;176:eA220019. doi:10.7326/A22-0019 [3]. This post is part of a series on crystal arthritis and osteoarthritis for interns and residents in internal medicine.

References

[1] Heijman MWJ, Fiolet ATL, Mosterd A, et al. Association of Low-Dose Colchicine With Incidence of Knee and Hip Replacements: Exploratory Analyses From a Randomized, Controlled, Double-Blind Trial. Ann Intern Med. 2023;176. doi:10.7326/M23-0289

[2] Schieker M, Conaghan PG, Mindeholm L, et al. Effects of Interleukin-1β Inhibition on Incident Hip and Knee Replacement: Exploratory Analyses From a Randomized, Double-Blind, Placebo-Controlled Trial. Ann Intern Med. 2020;173:509-515. doi:10.7326/M20-0527

[3] Centor RM, Neogi T. Annals On Call - Low-Dose Colchicine and Incident Knee and Hip Replacements. Ann Intern Med. 2023;176:eA220019. doi:10.7326/A22-0019

Gout gets all the attention. It has a famous name, a famous crystal, and a treatment algorithm every internist can recite. Calcium pyrophosphate deposition disease, still widely known by its older name, pseudogout, gets far less, even though Dr. McLean’s clinical impression is that it may be just as common in older patients. If you are waiting for a hyperuricemic patient with a red, hot big toe before you think about crystal arthritis, you will miss an entire disease that behaves differently, strikes different joints, and needs a different diagnostic mindset.

My guest was Dr. Robert McLean, Associate Clinical Professor of Medicine at Yale School of Medicine, who also practices rheumatology and is President Emeritus of the American College of Physicians [4]. He has taught this topic to residents and fellows for years, and I wanted him to walk our listeners through it the way he walks his trainees through it.

A Diagnosis Born in the Annals

Gout crystals were first identified under polarizing microscopy around 1960 to 1961. In May 1962, not 1963 as recalled in our conversation, Daniel McCarty and colleagues published a case series in Annals of Internal Medicine describing eight patients whose joints looked like gout clinically, except the pattern was wrong: the knee, the wrist, and the elbow, not the base of the great toe [1]. Under polarized light, most of these patients showed a different crystal altogether. A companion paper published the same month worked out its chemical identity: calcium pyrophosphate dihydrate [2]. Fifteen years later, in a short 1977 editorial, McCarty himself named the condition and proposed diagnostic criteria for it, coining the term “pseudogout” [3]. Dr. McLean recalled this second piece as dating to 1973 and focusing mainly on pathophysiology. Checking the primary source: it is a 1977 article, and it reads more as a nomenclature and diagnostic-criteria piece than a mechanistic one, though it does touch on the biology McCarty and colleagues had been building toward for a decade. The discovery happened entirely within living memory of physicians still practicing today, which is part of why the field feels comparatively young and unsettled next to gout.

Recognizing the Pattern

The classic presentation Dr. McLean teaches: an elderly woman with an acutely hot, swollen wrist and nothing else going on. Why women more than men, and why the wrist specifically, is not well understood. Chondrocalcinosis, calcification within cartilage itself rather than bone, is the radiographic signature, and it shows up most predictably in two places: the triangular fibrocartilage complex of the wrist and the menisci of the knee. Guideline recommendations now support imaging beyond the acutely affected joint: if CPPD is on the differential, x-ray the wrist and knee even when the hot joint is elsewhere, since chondrocalcinosis found incidentally at one of these classic sites strengthens the case. An x-ray cannot rule out infection and cannot rule in or out early erosive rheumatoid arthritis, so it is a supporting clue, not a substitute for fluid analysis.

A normal x-ray or absent chondrocalcinosis does not exclude CPPD, and calcification without an acute joint does not confirm it. Chondrocalcinosis is common and frequently asymptomatic in older adults; it tells you deposition has occurred, not that it is the cause of today’s hot joint.

Tests to Consider

The joint still has to be tapped. Classic teaching required seeing a crystal actually phagocytosed inside a polymorphonuclear cell before attributing the inflammation to that crystal, because patients with hyperuricemia can have incidental urate crystals sitting in a joint that is actually inflamed by something else entirely, CPPD included. In McCarty’s original series, one patient had both crystal types present. The corollary that matters practically: a first hot joint, especially in a patient who could plausibly have either infection or crystal disease, gets a cell count, a crystal search, a Gram stain, and a culture. You cannot call it clean until the culture comes back negative, particularly in patients with diabetes or other reasons to worry about a blunted immune response.

A practical caveat for trainees outside academic or rheumatology-staffed settings: reliable polarized-light crystal identification is not universally available. Send the fluid anyway for cell count, Gram stain, and culture, and get crystal analysis wherever it can be arranged, even if the result lags; a delayed or unavailable crystal read does not remove the need to rule out infection in the meantime, and a hot joint should not be managed as though the diagnosis is settled before that.

Two newer imaging tools have a role, mostly for gout rather than CPPD. Dual-energy CT is now well validated for identifying urate deposits and has largely a supportive, less well-established role for CPPD specifically. Ultrasound’s double-contour sign is specifically a gout finding; ultrasound-based CPPD criteria exist in the research literature but were not part of this conversation and are not yet standard clinical practice. Neither imaging tool replaces synovial fluid analysis as the diagnostic anchor.

For clinicians who want a systematic reference point rather than pattern recognition alone, the 2023 ACR/EULAR classification criteria formalize much of what Dr. McLean describes intuitively: a history of joint pain, swelling, or tenderness plus either crowned dens syndrome or CPP crystals on synovial fluid analysis is sufficient by itself, and short of that, a weighted point system combining clinical features, associated metabolic disease, and imaging can classify a patient [5]. These are classification criteria built for research cohorts, not diagnostic criteria for an individual patient in front of you, but they are a useful checklist for what actually distinguishes CPPD from its mimics.

Treatment: Familiar Drugs, Familiar Caveats

Once CPPD is confirmed or strongly suspected, treatment tracks closely with gout. Colchicine, started early, appears most effective: three times on day one, then twice daily thereafter, with reassessment within a day or two. If response is incomplete, an NSAID or a short course of steroids, oral or as a joint injection, is added. For a patient with recurrent, previously characterized pseudogout who looks otherwise well, with clear yellow, non-turbid fluid, it is reasonable to aspirate and inject steroid at the same visit rather than waiting for fluid results, provided infection risk is genuinely low.

Colchicine’s real hazards are narrower than the reflexive caution around it in chronic kidney disease suggests. It does not cause kidney dysfunction directly, but it accumulates in renal impairment and should be dosed less frequently, not avoided outright. Bone marrow suppression is a real but largely long-term-use concern, most relevant with pre-existing marrow disease such as leukopenia. Dr. McLean’s own practice is to dose-adjust colchicine in CKD rather than withhold it, a distinction worth making explicit rather than defaulting to avoidance.

For patients who cannot take colchicine, NSAIDs, or systemic steroids, whether from advanced CKD, diabetes, or another contraindication, IL-1 inhibition (anakinra, or the longer-acting canakinumab) is an option Dr. McLean raised. This deserves an explicit flag: canakinumab received FDA approval in 2023 for gout flares in patients who cannot use NSAIDs, colchicine, or repeated corticosteroid courses, but neither anakinra nor canakinumab carries an FDA indication for CPPD specifically. Extending either drug to CPPD rests on the shared IL-1-driven, crystal-induced inflammatory mechanism the two diseases have in common, not on CPPD-specific trial data. A local steroid injection remains the more conventional fallback when systemic options are unavailable.

The Overlap With Osteoarthritis

One of the more interesting threads in this conversation concerns chronic, low-grade CPPD and its relationship to osteoarthritis. Osteoarthritis is increasingly understood as a less inflammatory disease rather than a non-inflammatory one, and Dr. McLean’s view is that a subset of atypical osteoarthritis, particularly osteoarthritis concentrated at the second and third metacarpophalangeal joints, a distribution osteoarthritis does not usually pick, may represent smoldering CPPD. An unusual distribution of osteoarthritis is a pattern-recognition cue to look for a secondary cause, not just a variant of ordinary wear and tear. The secondary causes worth considering are hemochromatosis, hyperparathyroidism, and hypomagnesemia; the last is worth a specific mention because it turns up in patients on calcineurin inhibitors such as tacrolimus after solid organ transplant, and can present as recurrent, low-grade, smoldering crystal disease rather than a dramatic acute flare. How aggressively to screen for these, including whether to check a parathyroid hormone level in a patient with a normal serum calcium, is judgment rather than protocol; Dr. McLean described a low threshold to do so rather than a fixed rule.

Crowned Dens Syndrome: The Case You Cannot Aspirate

CPPD can deposit at the atlantoaxial joint, producing crowned dens syndrome: severe neck pain from CPP crystal deposition around the odontoid process, visible on CT. The 2023 classification criteria list it, alongside synovial fluid crystals, as one of two findings sufficient by itself to classify a research subject as having CPPD [5], a formal echo of the pattern-recognition call rheumatologists were already making. This is a joint that cannot be tapped, which makes the diagnosis and the treatment decision unusually uncomfortable. Dr. McLean described managing one such case with a month of low-dose prednisone followed by a taper, without the reassurance a synovial fluid result would normally provide. When the joint in question cannot be sampled, the diagnosis rests on imaging and clinical pattern alone, and empiric treatment carries more uncertainty than it does everywhere else in this disease; a classification criterion built for research cohorts is a useful anchor for that judgment, not a substitute for it.

Prevention: The Honest Answer Is We Do Not Know

Gout has urate-lowering therapy. CPPD has nothing analogous, and Dr. McLean was candid that this is a real frustration given how much has been learned about the disease’s biology since 1962. For patients with recurrent attacks, low-dose colchicine, daily or every other day, is used empirically by many rheumatologists, but there is no trial establishing that it prevents future episodes. The absence of an evidence base for CPPD prophylaxis is worth stating plainly rather than papering over with an extrapolated gout regimen.

Bob’s Pearls

1. Chondrocalcinosis on x-ray, especially at the wrist or knee, is the classic clue to CPPD, but the diagnosis still requires finding crystals in synovial fluid. Imaging supports; it does not substitute.

2. The textbook presentation is an elderly woman with an acute, isolated hot wrist and nothing else. It is worth having this pattern loaded before the patient walks in, because CPPD does not announce itself with a metabolic clue the way gout often does.

3. Colchicine works for CPPD much as it does for gout: three doses on day one, then twice daily, reassessed within 48 hours. In chronic kidney disease, dose-interval adjustment, not avoidance, is the right response; the drug does not cause direct kidney injury, it accumulates.

4. Osteoarthritis in an atypical distribution, particularly the second and third MCP joints, should prompt a look for CPPD and its secondary causes: hemochromatosis, hyperparathyroidism, and hypomagnesemia, the last worth remembering specifically in transplant patients on tacrolimus.

5. Unlike gout, CPPD has no proven preventive therapy. Empiric low-dose colchicine is common practice, not evidence-based prevention, and trainees should be able to tell a patient the difference.

6. Crowned dens syndrome: when the joint cannot be aspirated, diagnosis and treatment rest on imaging and clinical judgment alone. The 2023 ACR/EULAR classification criteria list this presentation as sufficient to classify a research subject, formalizing what experienced rheumatologists already do by pattern recognition; it is a research tool that echoes the clinical judgment, not a license that replaces it.

This podcast episode featured Dr. Robert M. McLean, Associate Clinical Professor of Medicine at Yale School of Medicine and President Emeritus of the American College of Physicians. The articles discussed are: McCarty DJ, Kohn NN, Faires JS. “The Significance of Calcium Phosphate Crystals in the Synovial Fluid of Arthritic Patients: The ‘Pseudogout Syndrome’: I. Clinical Aspects.” Ann Intern Med. 1962;56:711-737 [1]. Kohn NN, Hughes RE, McCarty DJ, Faires JS. “...II. Identification of Crystals.” Ann Intern Med. 1962;56:738-745 [2]. McCarty DJ. “Calcium Pyrophosphate Dihydrate Crystal Deposition Disease: Nomenclature and Diagnostic Criteria” [editorial]. Ann Intern Med. 1977;87:240-242 [3]. Centor RM, McLean RM. “Annals On Call - Pseudogout Syndrome.” Ann Intern Med. 2025;178:e2502206OC [4]. Abhishek A, et al. “The 2023 ACR/EULAR Classification Criteria for Calcium Pyrophosphate Deposition Disease.” Ann Rheum Dis. 2023;82:1248-1257 [5]. This post is part of a series on crystal arthritis for interns and residents in internal medicine.

References

[1] McCarty DJ, Kohn NN, Faires JS. The Significance of Calcium Phosphate Crystals in the Synovial Fluid of Arthritic Patients: The “Pseudogout Syndrome”: I. Clinical Aspects. Ann Intern Med. 1962;56:711-737. doi:10.7326/0003-4819-56-5-711

[2] Kohn NN, Hughes RE, McCarty DJ, Faires JS. The Significance of Calcium Phosphate Crystals in the Synovial Fluid of Arthritic Patients: The “Pseudogout Syndrome”: II. Identification of Crystals. Ann Intern Med. 1962;56:738-745. doi:10.7326/0003-4819-56-5-738

[3] McCarty DJ. Calcium Pyrophosphate Dihydrate Crystal Deposition Disease: Nomenclature and Diagnostic Criteria [editorial]. Ann Intern Med. 1977;87:240-242. doi:10.7326/0003-4819-87-2-240

[4] Centor RM, McLean RM. Annals On Call - Pseudogout Syndrome. Ann Intern Med. 2025;178:e2502206OC. doi:10.7326/ANNALS-25-02206-OC

[5] Abhishek A, Tedeschi SK, Pascart T, et al. The 2023 ACR/EULAR Classification Criteria for Calcium Pyrophosphate Deposition Disease. Ann Rheum Dis. 2023;82:1248-1257. doi:10.1136/ard-2023-224575

In the previous post, I talked with Dr. Alexander Leung about his research showing that the seated saline suppression test — the most commonly used confirmatory test for primary aldosteronism — performs no better than chance. If you have not read that one yet, it is worth going back to it first. This post builds directly on those findings, but from a different vantage point.

Dr. Jordana Cohen is a nephrologist and clinical researcher at the University of Pennsylvania, where she specializes in complex hypertension, secondary hypertension, and chronic kidney disease. She wrote the editorial accompanying Leung’s study in Annals of Internal Medicine, titled “Rethinking Confirmatory Testing in Primary Aldosteronism.” I wanted to bring her in not only to discuss the Leung paper but to do something we did not have space for in that episode: ask what a nephrologist who runs a complex hypertension clinic actually does with all of this information on a daily basis.

The opening exchange in our conversation captures the headline well. I asked about her screening philosophy, and she said: “My approach is screen everybody. Because we’re underscreening so badly — we found in our study published in Annals a few years ago that less than 2 percent of appropriate people were being screened.”

That is where this episode starts — and where most of us still are in practice.

The Leung Study Through Dr. Cohen’s Eyes

Before getting to the practical implications, I asked Dr. Cohen to walk through what made the Leung study methodologically distinctive. Her answer was crisp.

Prior work on confirmatory testing for PA — and there have been 55 such studies — shared a fundamental flaw: they assessed the performance of confirmatory tests against other confirmatory tests, or against reference standards that only verified positive cases. In medicine, when a test says “positive,” we tend to follow up. When it says “negative,” we tend to stop. That means the false-negative rate of the index test is never measured, because nobody verifies the cases it said were negative.

What Leung’s group did differently was proceed to adrenal vein sampling and treatment in every patient, regardless of the saline suppression result — and do so blinded to it. Adrenal vein sampling told the clinicians whether disease was unilateral (surgical candidates) or bilateral (medical therapy). Treatment response — using the PASO criteria, assessed by physicians who had no idea what the confirmatory test had shown — became the reference standard. If you had PA, you should respond to directed treatment. If you did not have PA, you should not.

The result, as Dr. Cohen put it: the confirmatory test “did a horrific job.” It missed a large number of patients who subsequently responded to treatment, and the overall discrimination was slightly better than rolling dice. She was careful to note one important point about the cohort’s composition: these were patients with high pretest probability — resistant hypertension, hypokalemia, or an adrenal nodule — and nearly all of them turned out to have PA. Critics have observed that this enriched prevalence inflates false-negative rates, which is arithmetically correct. But Dr. Cohen pointed out that the sensitivity analysis restricted to patients who lateralized and underwent surgery showed essentially the same result, and surgical response is among the most objective outcomes available.

She also anticipated an objection I raised from our earlier aldosteronism episodes: wasn’t it already known that patients with resistant hypertension tend to respond well to spironolactone, regardless of whether they have confirmed PA? She agreed — the PATHWAY-2 and ASCOT trials established exactly this. But the Leung cohort was not limited to resistant hypertension. It included patients with hypertension and an adrenal nodule or hypokalemia who did not have resistant hypertension. Those patients, she argued, would not be expected to respond well to spironolactone simply by virtue of having resistant blood pressure — so their response is harder to dismiss as nonspecific.

The Practical Case Against Confirmatory Testing

Dr. Cohen’s editorial takes a clear position: routine confirmatory testing should be abandoned. In our conversation, she expanded on why she has already abandoned it in her own clinic — and has for years.

Her clinic’s approach is to screen with renin and aldosterone, interpret the result, and proceed directly to either adrenal vein sampling or empirical mineralocorticoid receptor antagonist therapy. The confirmatory step is reserved for borderline cases — situations where the ARR is just below threshold and the clinical question is genuinely uncertain, and where the answer would materially change management. For patients with a clear picture — suppressed renin, elevated aldosterone, hypertension that fits — the confirmatory test adds nothing useful and may prevent patients from receiving treatment they need.

She was careful to name what is lost by skipping confirmatory testing: the small number of patients who are genuinely false positives on the ARR, for whom a true confirmatory step might prevent unnecessary workup or treatment. But given the poor specificity of the saline suppression test (21 to 57 percent across cutoffs in the Leung data), this protection is largely illusory. A negative confirmatory test does not reliably identify true negatives. What it does, in practice, is create a reason to stop — and stopping is what has kept the diagnosis rate below 2 percent.

A practical note for trainees: adrenal vein sampling requires specialized expertise and is not reliably available at most hospitals. For patients who screen positive and appear to be potential surgical candidates, the right move is referral — to an endocrinologist or a nephrologist with a specific interest in hypertension — rather than attempting to navigate the AVS pathway locally with a team that does it rarely. For patients who are not surgical candidates, empirical MRA therapy is the appropriate next step and does not require AVS at all.

Screening: Who, and How

I asked Dr. Cohen to sit down, figuratively, with a group of primary care physicians and tell them who to screen and when.

She described her own approach as “a little bit controversial” — and then named it: screen everybody with hypertension. Not just resistant hypertension, not just patients with hypokalemia or an adrenal nodule, but anyone with elevated blood pressure. She grounded this in several lines of evidence. Her own prior study in Annals of Internal Medicine documented the 2 percent screening rate. Anand Bhatia’s group has consistently estimated that 10 to 20 percent of patients with hypertension have primary aldosteronism. And work from Greg Hundemer — one of the co-authors of the Leung paper — has made the case that aldosterone is directly toxic to the kidney and cardiovascular system: autonomous aldosterone production is likely a major contributor to what we call hypertensive nephropathy, and probably contributes to heart failure as well. If that evidence is right, the cost of missing the diagnosis early is organ damage that accumulates silently while blood pressure is being controlled with the wrong drug.

She noted that the 2024 European Society of Cardiology guideline now recommends screening anyone with a confirmed blood pressure above 140/90, regardless of whether other traditional risk factors for PA are present. She expected other major guidelines to follow.

US guidelines have not yet adopted this broad a threshold, and Dr. Cohen’s “screen everybody” approach is ahead of current American consensus. A practical framework for US trainees that is both evidence-based and defensible: screen all patients with resistant hypertension, screen all patients with hypertension and hypokalemia — even when the hypokalemia is attributed to a thiazide or loop diuretic (the diuretic may be unmasking underlying PA, not causing the low potassium), and screen anyone with difficult-to-control hypertension before they have formally met the resistant hypertension definition. This captures the highest-yield population without requiring a posture that attending physicians are not yet taking.

How to screen: check a plasma renin activity and plasma aldosterone. That is it. No medication holds, no special timing, no position requirements. She mentioned preferring renin activity over direct renin concentration because renin activity uses a consistent threshold across all laboratories, while direct renin thresholds vary by lab — a practical point that will resonate for anyone who has tried to interpret a direct renin result from an unfamiliar reference range. If the renin is suppressed and the aldosterone is elevated, the conversation has started. If the picture is clear, proceed from there. If the result is borderline, you can always repeat it.

CT Scan: For One Reason Only

I asked Dr. Cohen when she orders a CT scan, given our earlier discussions about the biology of primary aldosteronism — the fact that most patients have diffuse or multifocal autonomous production rather than discrete adenomas, and that imaging finds incidentalomas that are often non-functioning.

Her answer was precise. She orders a one-time CT scan in every patient she evaluates for PA — but for a specific reason: to exclude adrenal carcinoma. Aldosterone-secreting adrenocortical carcinoma is rare but real, and you do not want to send a patient with an adrenal malignancy for endocrine management of hypertension.

Beyond that, she does not use the CT to determine treatment. And her reasoning is worth absorbing. Adrenal adenomas are extremely common — she compared them to kidney cysts, the kind of thing you find more often with age. Many are non-functioning. A patient can have an adenoma on one side and still have bilateral disease driven by the other adrenal gland, or by diffuse hyperplasia. And conversely, a patient can have unilateral disease that lateralizes clearly on adrenal vein sampling, with no visible adenoma on CT. She cited cases from her center where the adenoma visible on imaging was a red herring — on the wrong side — while the true source of autonomous production was found on the opposite gland only by AVS. CT does not replace adrenal vein sampling, and it should not be used to decide whether to pursue it.

This connects directly to the theme of the previous post, where Dr. Leung described how the complexity and length of the traditional diagnostic pathway — screen, confirm, image, sample, then treat — is itself one of the reasons PA stays underdiagnosed. Every step is an opportunity to stop. Clarifying what CT does and does not contribute is part of streamlining that pathway: one scan for one reason, then move on.

Mineralocorticoid Receptor Antagonists: Earlier and More Often

One thread I pursued was whether MRAs should be reaching patients earlier in the cascade — not just as the fourth drug in resistant hypertension, but as a second or third agent for patients found to have PA earlier in their hypertensive course.

Dr. Cohen confirmed this is already how she practices. She is using spironolactone and eplerenone earlier — not because of a guideline change, but because she is screening more broadly and finding PA in patients who are not yet at the resistant hypertension threshold. When you find the disease in someone with stage 1 or stage 2 hypertension, there is no reason to work through three other antihypertensives before reaching the drug that addresses the underlying pathophysiology.

For men, she tends to prefer eplerenone over spironolactone because of the gynecomastia risk that comes with spironolactone’s anti-androgenic effects. For younger women, spironolactone has actually gained traction in primary care, partly because of its use in polycystic ovarian syndrome and acne — she sees a number of referrals from primary care physicians who started spironolactone for a dermatologic or endocrine indication and noticed a blood pressure benefit.

She also raised the future landscape. Aldosterone synthase inhibitors — a newer class that directly blocks the enzyme producing aldosterone — are entering clinical practice. The lorandostrat trial, published recently in the New England Journal of Medicine, used the drug as a third agent in patients with resistant hypertension, and the early results were promising. These are not yet FDA-approved for hypertension, and approval for a broader indication will require further study — but they represent a potentially meaningful addition to the therapeutic options available for autonomous aldosterone excess.

Amiloride: The Underused Option

Toward the end of our conversation, Dr. Cohen raised a drug that does not often appear in aldosteronism discussions: amiloride. This is a potassium-sparing diuretic that works by directly blocking the epithelial sodium channel (ENaC) in the renal tubule — the channel that aldosterone normally activates.

A recent trial published in JAMA compared amiloride to spironolactone for resistant hypertension in a randomized design and found them to perform similarly. This did not surprise her. She and other hypertension specialists had long used amiloride empirically, particularly in patients who had a suppressed renin with a normal or low aldosterone — a pattern that suggests not classic PA, but rather a Liddle syndrome-like physiology, where the ENaC channel is tonically overactive without requiring elevated aldosterone. Amiloride targets that channel directly. For those patients, it works well.

Two practical advantages she mentioned: amiloride does not cause gynecomastia, and it is inexpensive. It is, she noted with some amusement, a drug that frequently generates pharmacy calls asking whether the prescriber meant amlodipine — a confusion her clinic has become very accustomed to fielding.

The finding from the JAMA trial — that the renin level did not predict response to amiloride, contrary to what many had assumed — was new information for her, and she expects it will expand her use of the drug as a fourth-line agent even in patients without the classic low-renin, low-aldosterone phenotype. This sits in some tension with the clinical intuition that amiloride is especially suited to the low-renin, low-aldosterone pattern — if renin level does not predict response, the rationale for selecting amiloride on that basis becomes less clear. Whether renin level should guide the choice between amiloride and an MRA remains an open question. More data are needed.

Bob’s Pearls

1. Screen resistant hypertension, hypertension with hypokalemia, and difficult-to-control hypertension — do not wait for US guidelines to catch up. Less than 2 percent of eligible patients are being tested, and mounting evidence suggests 10 to 20 percent of patients with hypertension have some form of autonomous aldosterone production. Dr. Cohen screens everybody; the 2024 ESC guideline supports her. A practical starting point for US trainees: all resistant hypertension, all hypertension plus hypokalemia regardless of whether a diuretic is blamed (the diuretic may be unmasking PA, not causing the low potassium), and anyone whose blood pressure is proving difficult to control before they formally meet the resistant definition. Whatever your threshold, the gap between what we should be doing and what we are doing is enormous.

2. Checking renin and aldosterone is all you need for screening — and it is simpler than it sounds. Plasma renin activity and plasma aldosterone, drawn without medication holds or special preparation. Renin activity has a consistent threshold across laboratories. If the renin is suppressed and the aldosterone is elevated, the conversation has started.

3. The confirmatory test is probably more harmful than helpful. Dr. Cohen has not used confirmatory testing routinely in her clinic for years. After a clear positive screen, she proceeds directly to adrenal vein sampling in surgical candidates and empirical MRA therapy in those who are not. The Leung data support this approach — a test that performs at chance adds friction without adding information.

4. CT scanning has a role, but a narrow one. One CT scan to exclude adrenal carcinoma is reasonable. Beyond that, the scan does not determine treatment and should not determine whether adrenal vein sampling is pursued. Adenomas can be non-functioning red herrings; unilateral disease can exist without a visible adenoma; and the opposite gland can be the true culprit even when a nodule is present.

5. Think about aldosterone’s role in kidney and heart disease — not just blood pressure. Autonomous aldosterone production is likely a major contributor to hypertensive nephropathy and may contribute to heart failure. When we treat blood pressure with the wrong drug and leave aldosterone excess unaddressed, we may be controlling the number while allowing the organ damage to continue. Finding and treating PA early may matter for reasons that go well beyond blood pressure targets.

This podcast episode featured Dr. Jordana Cohen, nephrologist and clinical researcher at the University of Pennsylvania. The articles discussed are: Leung AA et al., “Confirmatory Testing for Primary Aldosteronism: A Study of Diagnostic Test Accuracy,” Annals of Internal Medicine, 2025 (doi:10.7326/ANNALS-24-03153; corrected June 2026); and Cohen JB, “Rethinking Confirmatory Testing in Primary Aldosteronism,” Annals of Internal Medicine, 2025 (doi:10.7326/ANNALS-25-01368). This post is part of a series on primary aldosteronism for interns and residents in internal medicine.

A note on sequencing: this episode aired in December 2020 — several years before the Leung study on confirmatory testing that appeared in the previous post. I am presenting it here because it provides the epidemiologic foundation that makes the Leung findings matter. Knowing that the confirmatory test does not work is important. Knowing how far we are from even reaching the confirmatory test — how many patients with primary aldosteronism never get screened in the first place — is what makes the whole diagnostic pathway question urgent. This is precisely why I keep making podcasts and writing Substack posts about primary aldosteronism. Primary aldosteronism is a rapidly evolving field, and many physicians currently in practice were trained before the concepts laid out in this series were well established. The gaps documented in this study are not simply a failure of attention — they also reflect a generation of clinicians who were taught a different framework. Closing that gap is the whole point of this series.

My guest was Dr. Jordana Cohen, who at the time was an assistant professor of medicine and epidemiology at the University of Pennsylvania, focused on complex hypertension, secondary hypertension, and chronic kidney disease — and who describes herself, accurately, as a data nerd. Regular listeners will recognize her name from the next episode in this series, where she returns as the author of the editorial accompanying the Leung confirmatory testing paper. But this episode came first, and it stands on its own.

The article we discussed: “Testing for Primary Aldosteronism and Mineralocorticoid Receptor Antagonist Use Among U.S. Veterans, a Retrospective Cohort Study,” published in Annals of Internal Medicine on December 29, 2020.

The Question Behind the Study

Guidelines have recommended screening patients with resistant hypertension for primary aldosteronism for many years. The evidence base for why to screen is solid: PA is common in this population, the cardiovascular risk it confers goes beyond blood pressure, and effective treatments exist. The recommendation has been there. The question Dr. Cohen and her colleagues wanted to answer was simpler and more uncomfortable: are physicians actually doing it?

Prior studies at individual health systems and referral centers had suggested the answer was no — that only 2 to 3 percent of eligible patients were being tested. But those studies were done in settings where patients receive their primary care in one place and specialty care somewhere else entirely. It was possible that testing was happening and just not being captured in any single system’s records. Dr. Cohen wanted to close that loophole.

The VA was the right place to do it. The VA is a closed system: patients who use the VA for their primary care tend to use it for most of their other care as well. There are no gaps between where the testing gets ordered and where the result gets recorded. And the VA’s pharmacy fill data allowed Dr. Cohen to know not just what was prescribed, but what patients actually filled — a meaningful proxy for adherence that is rarely available in other databases.

The Cohort

The study included approximately 270,000 veterans who developed new-onset resistant hypertension — defined as hypertension requiring at least three antihypertensive medications, or requiring a fourth drug to reach control — while being followed in the VA system. It had to be incident resistant hypertension, not prevalent: patients had to have developed it during the observation period, so the investigators could study what happened in response to it.

The cohort reflects the VA population: median age 65, 4 percent women — a small percentage, but still more than 11,000 women — and 19 percent Black non-Hispanic, representing roughly 51,000 Black veterans. About half the cohort was obese. Only 3 percent had hypokalemia, defined as a potassium of 3.5 or below, at any point before the index date. That last number is worth pausing on: if you are still waiting for low potassium before you screen for PA, you are going to miss 97 percent of the people in this cohort who might have it.

What They Found

Of 270,000 patients with resistant hypertension, approximately 4,000 were tested for primary aldosteronism — about 1.5 percent, consistent with the lower end of what prior smaller studies had shown. The rate had not improved over time. It did not vary meaningfully by center size or by whether the center was academic or community-based. It was uniformly, depressingly low everywhere.

The treatment picture was similarly stark. Only 13 percent of patients with resistant hypertension were receiving a mineralocorticoid receptor antagonist. Given the evidence from PATHWAY-2 and ASCOT that spironolactone should be the fourth drug in resistant hypertension, this is a gap that cannot be explained by clinical uncertainty. The drug works. The recommendation has been in guidelines for years. Thirteen percent.

The investigators also looked at the relationship between testing and treatment. Patients who were tested for primary aldosteronism were four times more likely to be started on a mineralocorticoid receptor antagonist than those who were not. This makes sense: testing focuses clinical attention on the aldosterone pathway. It is not that testing caused the MRA — it is that clinicians who thought hard enough about PA to order the test also thought hard enough to prescribe the appropriate drug.

Two other findings deserve mention. Among the provider and center-level factors the investigators examined, specialty made a large difference: endocrinologists and nephrologists were far more likely to test than primary care physicians or cardiologists. And rural centers were less likely to test than urban ones — an important access disparity that points toward a real implementation challenge in getting PA recognized outside of academic or specialty-dense settings.

The Blood Pressure Finding

One of the more striking secondary findings was what happened to blood pressure over time. Patients who were tested for primary aldosteronism started out with higher blood pressure than those who were not — about 5 to 6 mmHg higher at baseline. And yet, over approximately four years of follow-up, with a median of 17 blood pressure measurements per patient, they ended up with lower blood pressure.

Dr. Cohen’s interpretation was careful: this is not necessarily evidence that testing itself lowered blood pressure. It is more likely that the clinicians who were testing for PA were also managing hypertension more attentively and aggressively across the board. Testing for PA is a marker of engaged, systematic care — and that engagement carries over into the overall management of these patients. Still, the observation reinforces the same message from a different angle: thinking about PA leads to better outcomes for these patients.

The Treatment Barriers: Real and Imagined

I asked Dr. Cohen to speak to the barriers that keep clinicians from prescribing spironolactone, because the 13 percent treatment rate suggests that something beyond simple ignorance is at play.

Two concerns came up most often. The first is hyperkalemia. Spironolactone does raise potassium, and the hyperkalemia episodes that followed broader adoption of the drug in heart failure patients a generation ago left a lasting impression on many physicians. But Dr. Cohen’s point is worth holding onto: patients with advanced chronic kidney disease, who carry the highest hyperkalemia risk, were excluded from her cohort. For the remaining patients, starting low and titrating slowly — monitoring potassium at two weeks and again with any dose change — is feasible and generally safe. The fear has been calibrated to the worst cases, not to the typical patients who would benefit.

The second barrier is gynecomastia, and it was a particularly salient issue in a cohort that was 96 percent male. Spironolactone’s anti-androgenic effects can cause breast tissue development in men, and this was almost certainly part of why prescribing rates were so low in the VA population. The practical answer, which Dr. Cohen raised, is eplerenone — a selective mineralocorticoid receptor antagonist without the anti-androgenic side effects. Crucially, eplerenone is now available as a generic, which removes the cost barrier that once made it a difficult second-line choice. For male patients, eplerenone should be the default consideration.

There was also a brief exchange about amiloride — a potassium-sparing diuretic that works by blocking the epithelial sodium channel in the renal tubule. Dr. Cohen noted that amiloride is probably equivalent to MRA therapy in resistant hypertension in general, and that it represents a practical option when MRAs cannot be tolerated. The PATHWAY-2 mechanisms substudies, published in Lancet Diabetes & Endocrinology in 2018, documented this directly: amiloride reduced clinic systolic blood pressure by 20.4 mmHg compared with 18.3 mmHg for spironolactone — essentially equivalent — and the authors concluded that amiloride should be considered a substitute when spironolactone is not tolerated. The evidence base supporting this option has continued to grow since this podcast aired. For patients with confirmed primary aldosteronism, however, Dr. Cohen made an important distinction: MRA therapy may have advantages beyond blood pressure, because aldosterone acts on mineralocorticoid receptors throughout the cardiovascular system and kidney, and direct antagonism at those receptors may offer target organ protection that simply blocking the downstream sodium channel does not provide.

Testing in the Hospital

I raised a question that comes up for many hospitalists and residents: if a patient is admitted with severe or resistant hypertension and nobody has ever worked them up for PA, is it worth ordering the test in the hospital?

Dr. Cohen’s answer was qualified but useful. Renin is affected by posture — upright activity suppresses renin less than lying flat does, which matters for interpreting the result. Her recommendation was to check the test toward the end of the hospitalization, after the patient has been up and moving for an hour or two. That timing gives you a result that is more physiologically representative. It is not a perfect screening test in the hospital, but it is far better than sending the patient home without any data and hoping the outpatient team picks it up. Drawing renin and aldosterone, even imperfectly, plants the flag.

Bob’s Pearls

1. Two percent of eligible patients are being tested. That is the headline — and it has not improved. The VA study established this at scale, in a closed system where missing data is not the explanation. The problem is not that we lack a screening test. The problem is that we are not using it.

2. Thirteen percent of patients with resistant hypertension are receiving a mineralocorticoid receptor antagonist. Guidelines have recommended spironolactone as the fourth drug in resistant hypertension for years. The evidence from PATHWAY-2 and ASCOT is solid. The gap between what we know and what we prescribe is not a knowledge problem — it is a habit and attention problem.

3. Hypokalemia was present in only 3 percent of this cohort. If you are waiting for a low potassium to screen for primary aldosteronism, you will miss the vast majority of patients who have it. Hypokalemia is a late finding, not a prerequisite.

4. Eplerenone is now generic. Cost was once a real barrier to choosing eplerenone over spironolactone in men. That barrier is largely gone. Gynecomastia should no longer be a reason not to treat.

5. Testing in the hospital is reasonable — check toward the end of the stay. Renin is influenced by position and activity level. After the patient has been up and moving for an hour or two, drawing renin and aldosterone is a practical and useful starting point for the outpatient team.

6. When MRAs cannot be tolerated, amiloride is a well-supported alternative fourth drug. The PATHWAY-2 mechanisms substudies showed amiloride reduced systolic blood pressure essentially as well as spironolactone in resistant hypertension, and the authors explicitly proposed it as a substitute when spironolactone is not tolerated. Gynecomastia is a real barrier in a predominantly male population, and not every patient can use eplerenone — amiloride gives those patients a fourth drug backed by randomized data rather than a dead end. The evidence supporting this option has continued to accumulate since this podcast aired in 2020. The important caveat: in confirmed PA, mineralocorticoid receptor antagonism may offer target organ protection beyond blood pressure that amiloride, acting downstream on the sodium channel, may not fully replicate.

This podcast aired in December 2020. The article is: Cohen JB et al., “Testing for Primary Aldosteronism and Mineralocorticoid Receptor Antagonist Use Among U.S. Veterans, a Retrospective Cohort Study,” Annals of Internal Medicine, 2020. This post is part of a series on primary aldosteronism for interns and residents in internal medicine.

If you have been following this series, you know that I am somewhat obsessed with primary aldosteronism. The first post covered the basics: the disease is far more common than most clinicians were taught, hypokalemia is a late and unreliable sign, and the aldosterone-to-renin ratio is the recommended screening test. The second went deeper — into a study challenging whether the ARR is as sensitive as we thought, and proposing that renin suppression may be the more fundamental signal to track.

This episode takes the next step in the diagnostic pathway and asks a question that turns out to have a striking answer: once you have a positive screen, does the confirmatory test you are supposed to order actually work?

My guest was Dr. Alexander Leung, a clinical endocrinologist and associate professor at the University of Calgary. He is the lead author of the study we discussed: “Confirmatory Testing for Primary Aldosteronism: A Study of Diagnostic Test Accuracy,” published in Annals of Internal Medicine in 2025. The short answer to my question: no, the confirmatory test does not work — and Dr. Leung’s group is the first to have demonstrated this rigorously, by doing something that sounds obvious but apparently nobody had done before.

A word of context before we go further. Dr. Leung is not a disinterested observer here — he is the lead author of the study under discussion, and some of his conclusions about what should change in clinical practice may be ahead of where the broader field currently stands. I found his argument compelling and the data behind it hard to dismiss. But I also want you to be able to distinguish between what the study demonstrates and what Dr. Leung believes should follow from it. The next post in this series features an editorial on the same article by Dr. Jordana Cohen — a nephrologist and hypertension researcher at Penn — which will give you a second, independent perspective on what these findings mean in practice.

The Diagnostic Pathway — and Why It Intimidates Physicians

Dr. Leung opened by naming what he believes is the central reason primary aldosteronism is so dramatically underdiagnosed: the diagnostic pathway is simply too daunting.

Even experienced, conscientious internists and family physicians get intimidated when they consider working up a patient for PA. And it is not hard to see why. The traditional pathway, as it was codified in the 2016 Endocrine Society guidelines and reflected practice going back even further, involves multiple sequential steps — each one requiring expertise, coordination, and patience. You screen with the ARR. If that is abnormal, you confirm with a dynamic suppression test. If confirmation is positive, you refer for adrenal vein sampling to determine whether the excess aldosterone is coming from one gland or both. Then, depending on what AVS shows, you either refer for surgery or commit the patient to long-term medical therapy. As Dr. Leung put it, for the average frontline physician, the prospect of a patient with PA triggers thoughts like: “Oh, another person I have to put through multiple steps before I can even have a diagnosis, before I can even know whether I should be treating with some sort of targeted treatment.”

The confirmatory testing step is one of the friction points in that pathway. And as Dr. Leung has spent years studying, the evidence base underneath it is much shakier than anyone had bothered to examine carefully.

The Logic of Confirmatory Testing

The idea behind confirmatory testing is straightforward and physiologically sound — or at least it sounds that way. In endocrinology, when a hormone is too high, you try to suppress it. If it cannot be suppressed, that suggests autonomous production that is no longer under physiologic regulation. In the case of primary aldosteronism, the suppression challenge is sodium loading: give the patient a large salt load, which should suppress renin, which should in turn pull aldosterone down. If aldosterone stays elevated despite volume expansion, the conclusion is that something is producing it autonomously.

The most commonly used version of this test in North America became the seated saline suppression test — two liters of normal saline infused intravenously over four hours, with aldosterone measured at the end. The seated position was thought to improve accuracy compared to the traditional supine protocol. If aldosterone after infusion exceeded a certain threshold, you confirmed the diagnosis and proceeded. If it did not, you could, theoretically, rule PA out.

The problem — and Dr. Leung documented this exhaustively — is that the evidence base for this test was built on a methodological flaw so fundamental it should give all of us pause.

The Tautological Fallacy: Using a Test to Prove Itself

If you look at all 55 historical studies that have examined the diagnostic accuracy of confirmatory testing for PA, you will see something remarkable: in the vast majority of them, the reference standard used to determine whether the confirmatory test was correct was itself another confirmatory test. In other words, the test was being used to validate itself. Dr. Leung called this exactly what it is — a tautological fallacy. A circular argument.

He offered an analogy that listeners to this podcast will likely recognize. For those of a certain era, PIOPED — or as Dr. Leung called it, PALPE2 — was a landmark study of VQ scanning for pulmonary embolism, using pulmonary angiography as the gold standard. But there was a design problem: if the VQ scan was abnormal, they went ahead and did the angiography to verify it. If the VQ scan was low probability, they did not. This meant the false-negative rate of VQ scanning was never actually measured. Negative cases were presumed to be true negatives, because no one verified them.

The same thing happened in PA for decades. Patients with a positive confirmatory test went on to further workup — adrenal vein sampling, and eventually treatment, which provided a way to know whether the diagnosis was right. But patients with a negative confirmatory test were sent away, and by and large nobody followed them to find out whether they actually had PA that was just missed. The false-negative rate of the confirmatory test was effectively unknown.

Dr. Leung and his group decided to find out.

The Study: Verifying Everyone, Regardless of Test Result

The key methodological innovation in the Calgary study was simple and, in retrospect, obvious: they took every patient who underwent the saline suppression test and proceeded to the next step in the pathway regardless of the result. The investigators were blinded to the saline suppression outcome. Patients with lateralizing disease on adrenal vein sampling went to surgery. Patients without lateralization received medical therapy — primarily spironolactone, with some eplerenone. And then treatment response was measured.

This matters because treatment response is the closest thing available to a real-world reference standard for PA. It is not a perfect gold standard — you cannot do bilateral adrenalectomies on everyone for pathological confirmation. But if a patient has PA and responds to directed treatment (surgical removal of a unilateral source, or blockade of the mineralocorticoid receptor), they meet what the authors defined using the PASO criteria: biochemical and clinical success. If they do not have PA, treatment should not work. This is what Dr. Leung called a “silver standard” — imperfect, but the best available.

One additional feature of the Calgary study that deserves mention: the center has a 96 percent success rate for adrenal vein cannulation, verified biochemically, sustained over more than two decades. That number is striking. In most centers around the world, a success rate of 70 percent is considered excellent; in many centers, the figure is 30 to 50 percent. The Calgary group’s ability to successfully sample both adrenal veins in virtually every patient is what made it feasible to universally verify the cohort.

What They Found

The results were stark.

When the saline suppression test result was unblinded and compared against treatment response, the test simply did not work. The most striking finding was the false-negative rate: at the most commonly used aldosterone cutoff, roughly 40 percent of patients who would go on to respond to PA-directed treatment had an aldosterone level that the test called “suppressed” — meaning the test told them they did not have significant PA when they did. Even at the lowest cutoff the investigators tested, 7 percent of true responders were still missed. The test performed similarly regardless of which aldosterone assay was used.

Many patients whose saline suppression test said “no PA here” — aldosterone suppressed appropriately — subsequently went on to respond to PA-directed treatment. They had the disease. The test said they did not.

I asked Dr. Leung directly about a scenario that strikes at the heart of the issue: what happened to patients who had an abnormal ARR, a negative saline suppression test, non-lateralizing disease on AVS, and were therefore put on spironolactone empirically? His answer was that more than 90 percent of those patients responded to treatment. Think about what that means in practice. Based on the aldosterone-to-renin ratio, these patients looked like they had PA. The confirmatory test said no, they don’t. And yet, when they were subsequently treated anyway, they got better.

Across every cutoff that has been proposed in the literature, across every subgroup analyzed, the result was the same: the saline suppression test did not distinguish patients who would respond to treatment from patients who would not.

A Note on What the Study Does and Does Not Show

The Calgary cohort is a specialized referral population. Every patient who walked into that clinic received an ARR — a practice Dr. Leung’s group has followed since 2005 — and patients were further selected as having a high enough pretest probability to proceed to adrenal vein sampling. The prevalence of confirmed PA in this cohort was approximately 91 percent.

When disease prevalence is that high, false-negative rates look worse, which is arithmetically expected. Critics of the study have made exactly this point. Dr. Leung acknowledged it directly in our conversation — and also pointed out that the subgroup of patients who lateralized and went to surgery showed a similar pattern, and that the surgical group itself is the most objective test of treatment response, since the adrenal gland is removed and histology is available.

The core finding — a 40 percent false-negative rate at the standard cutoff, consistent across assays and subgroup analyses — is robust. But the magnitude of the harm from false-negative tests in a lower-prevalence population, such as a primary care practice screening broadly, remains less precisely characterized.

What Dr. Leung’s Center Has Actually Done Since 2005

One of the most arresting moments of the conversation came when I asked Dr. Leung what his clinic had done in practice, aside from the research protocol. He told me that since 2005 — more than two decades — his center has not used confirmatory testing in clinical care. They screen with the ARR. They go directly to adrenal vein sampling in appropriate candidates. They treat. And in that time, they have successfully diagnosed and treated primary aldosteronism, including surgically curable unilateral disease, without a single confirmatory test in the clinical pathway.

That is not a research finding. That is a demonstration that the step can be skipped without apparent harm to diagnosis or outcomes.

He also noted that the 2025 Endocrine Society guidelines — which recommend considering PA screening in all adults with hypertension — were informed in part by this study. The landscape has already shifted.

What to Do When AVS Is Not Available

I pushed Dr. Leung on the practical reality that many clinicians — and most of the world’s hypertensive patients — do not have access to adrenal vein sampling. AVS requires an interventional radiology team with specific expertise in a technically demanding procedure. Dr. Leung noted that in most of Asia and South America, AVS is simply not available. Saying “screen and then refer for AVS” is advice that applies to only a fraction of the world.

His answer was direct: in settings without AVS, empirical mineralocorticoid receptor antagonist therapy — spironolactone is the most established agent — is the right approach. Treat based on the clinical picture and the ARR. Do not let the absence of AVS be a reason to stop after a positive screen and do nothing. The goal is to get treatment into the hands of patients who have autonomous aldosterone excess and are accumulating cardiovascular risk.

He added one important nuance. There is a narrow exception: a very young patient with a large adrenal nodule and extreme biochemical abnormalities may be a reasonable candidate to proceed directly to surgery without AVS, since the likelihood of unilateral disease is very high. But this is the tip of the iceberg — a small minority of PA patients. For the other 95 percent, do not let the lack of perfect workup infrastructure paralyze you.

Bob’s Pearls

1. The confirmatory test for primary aldosteronism has a false-negative rate too high to be clinically useful. At the standard aldosterone cutoff, 40 percent of patients who would respond to PA-directed treatment were called negative by the saline suppression test. This is the first study to rigorously verify every patient regardless of test result, finally measuring the false-negative rate that had been unknown for decades.

2. The evidence base for confirmatory testing was built on a tautological fallacy. Prior studies validated the test against itself, or against other confirmatory tests — a circular argument. Negative cases were presumed true negatives because nobody followed them to find out. The Calgary group is the first to have done so.

3. An abnormal ARR may be sufficient to act on — without a confirmatory step. Patients in the Calgary cohort who had a positive ARR, a negative saline suppression test, and non-lateralizing disease still responded to MRA therapy more than 90 percent of the time. The ARR, not the saline suppression test, was what predicted treatment response.

4. If AVS is not available, empirical MRA therapy is appropriate. Adrenal vein sampling requires specialized expertise and is simply not accessible in large parts of the world. For patients who screen positive and are not surgical candidates, or for whom AVS cannot be obtained, starting spironolactone is the right move. Not screening because you cannot complete the full pathway is the wrong move.

5. Do not let a negative confirmatory test close the case. If the clinical picture fits — suppressed renin, elevated aldosterone, hypertension that does not respond as expected — a negative saline suppression test should not reassure you. The test has now been shown to miss a large fraction of true cases. Screen, assess the ARR, and treat if the picture fits.

This podcast aired in August 2026. The article is: Leung AA et al., “Confirmatory Testing for Primary Aldosteronism: A Study of Diagnostic Test Accuracy,” Annals of Internal Medicine, 2025 (doi:10.7326/ANNALS-24-03153; corrected June 2026). The accompanying editorial is: Cohen JB, “Rethinking Confirmatory Testing in Primary Aldosteronism,” Annals of Internal Medicine, 2025 (doi:10.7326/ANNALS-25-01368). This post is part of a series on primary aldosteronism for interns and residents in internal medicine.

Bottom line up front: eGFR equations assume the creatinine has stopped moving. When it hasn’t, the calculated number doesn’t describe the patient’s current kidney function at all, it’s a snapshot of an unfinished process. And even once you’ve made that correction, a rising creatinine still isn’t a diagnosis. It’s a question: is this volume, or is this the kidney itself?

The case

A resident presents a patient on morning report. Creatinine 1.5, baseline 0.8. She reports the calculated eGFR along with it, the way we’re all trained to.

I stopped her there. A calculated eGFR on a creatinine that is actively rising is not a real number. It doesn’t tell you what the patient’s GFR is right now. It can’t.

Why the equation breaks down mid-rise

The eGFR equations, CKD-EPI included, are built on a steady-state assumption: creatinine production is constant, and at steady state, serum creatinine is inversely proportional to GFR. That relationship, production held in the numerator and clearance in the denominator, is what generates a stable, interpretable number when a patient’s kidney function isn’t changing.

But when creatinine is actively rising, the patient is, by definition, not at steady state. The serum level hasn’t caught up to whatever GFR is doing right now, it’s still climbing toward wherever it’s going to land. Plug that moving number into a steady-state equation and you get an answer that looks precise and means very little. The true GFR at that moment is very likely worse than even the higher creatinine suggests, because the creatinine is still behind the true drop in clearance, not ahead of it.

This is a distinct problem from the muscle-mass biases we covered in the first post in this series. That post was about a steady-state number being systematically wrong because of what’s feeding the numerator. This is about a number that was never at steady state to begin with. You cannot estimate a moving target with an equation built for a stationary one.

Here’s a thought experiment that shows why the rising number alone can’t tell you what you need to know. Imagine both kidneys are removed. GFR is now zero, as complete a kidney insult as exists. Creatinine production, though, hasn’t changed, it’s still being generated by muscle at the same fixed rate as before. With nowhere for it to go, that production rate alone determines how fast the serum level climbs. Starting from a baseline of 0.8, the creatinine the next day would land somewhere around 1.5 to 2.0, driven entirely by production with zero clearance.

That’s the point: a rise from 0.8 to 1.5 over a day or two looks, on paper, indistinguishable from complete, irreversible anephric failure, and it can just as easily represent something far milder and reversible. An increasing creatinine, by itself, does not carry enough information to tell you which of those you’re looking at. An eGFR calculated at that moment isn’t just biased, it’s answering a question the number was never built to answer, because the equation has no way to distinguish “no kidney function at all” from “kidney function that’s dropped some amount and creatinine hasn’t finished catching up.”

The bigger mistake: reaching for “AKI” too early

Here’s the trap that matters even more than the eGFR math. Once a creatinine is rising, it’s tempting to reach for a label: this is AKI. I’d go further: I think “AKI” itself is a problematic term, not because KDIGO’s creatinine and urine-output criteria are wrong as a trigger for concern, but because in practice the label gets used as an endpoint rather than a starting point. It sounds like a diagnosis, so it gets treated like one. My rule: don’t use the term “AKI” for a rising creatinine until you’ve considered volume and obstruction as possible causes. Reach for the label before doing that reasoning, and it papers over a question you haven’t actually answered yet: why is the creatinine rising?

At minimum, three very different processes can produce the identical number:

• Volume contraction. The kidneys are underfilled, GFR falls as a reversible, appropriate physiologic response, and the creatinine rises right along with it. Fix the volume, and the creatinine falls back down. Nothing was ever structurally wrong with the kidney.

• Obstruction. Something is blocking urine outflow, from the collecting system down to the urethra, and pressure backs up and drops GFR. This one is easy to miss because it’s mechanical rather than physiologic, and a bladder scan or renal ultrasound settles it quickly when it’s on the differential.

• A disease process actually injuring the kidney. Glomerular, tubular, interstitial, whatever the mechanism, this is real damage that volume repletion will not touch.

Same creatinine. Same eGFR, however unreliable. Completely different next step for each. Calling any of these “AKI” and moving on is exactly the kind of confident-sounding label that substitutes for diagnostic work rather than prompting it.

Obstruction deserves a place in the general differential every time a creatinine rises, but its pretest probability isn’t fixed. In a man in his 60s or 70s, prostatic enlargement makes obstruction one of the first things to rule out, often with nothing more than a bladder scan. This patient was a woman in her 60s, which makes obstruction meaningfully less likely as a first thought, though still worth a clean pass through the history (pelvic malignancy, prior pelvic surgery or radiation, retroperitoneal process) before setting it aside. That’s part of why volume and an intrinsic renal process carried more weight in the initial reasoning here.

Working through it at the bedside

Back to the patient. Blood pressure 110/60, off any blood pressure medications. Urinalysis: specific gravity 1.030, 2+ protein, 5 to 20 RBCs, 5 to 10 WBCs, several hyaline casts. Albumin 2.9.

This is a genuinely mixed picture, and it’s worth sitting with why, rather than skipping to the answer.

A specific gravity of 1.030 with a BP of 110/60 off medications is exactly what appropriately concentrating, underfilled kidneys look like. Volume contraction was a reasonable, well-supported hypothesis at this point, not a straw man to knock down. At the same time, the urinalysis has protein, red cells, white cells, and hyaline casts sitting alongside that concentrated specific gravity.

That protein result needed a second look before it could be trusted as evidence of anything. Dipstick protein is measured against urine concentration, and a specific gravity of 1.030 is concentrated enough that a 2+ reading could be a true finding or could be an artifact of how concentrated the urine is, dilute the same amount of protein into a less concentrated sample and the dipstick might read lower. Rather than treat 2+ protein at face value in either direction, the next step was an albumin to creatinine ratio, which corrects for urine concentration and gives an actual number to reason from instead of a semi-quantitative dipstick estimate.

With that picture, and low albumin that could point several directions, nutritional, inflammatory, or renal protein loss, three categories were on the differential from the start: malignancy, infection, and ANCA-associated disease, in addition to volume contraction. The active sediment didn’t get set aside in favor of volume, it stayed on the list while volume was tested, because the fluid challenge was cheap, low-risk, and worth doing regardless of what the sediment eventually turned out to mean.

Given a plausible, lower-risk, reversible explanation on the table, the reasonable move is to test it. That’s true regardless of what the sediment eventually turns out to mean, a BP of 110/60 is enough by itself to justify giving volume and seeing what happens, since the downside of trying is low and the potential payoff, correcting a reversible prerenal state, is high. Volume was repleted.

The creatinine kept rising anyway.

That result adds something important to the problem representation. It doesn’t just fail to help, it rules out volume contraction as the sole driver and reframes what needs explaining. Whatever is happening now needs an explanation that isn’t fluid status.

The reveal

The rest of the story had been sitting there the whole time. This patient had been admitted for fevers, sweats, and weight loss, significant, over two months, along with dyspnea and blood in her sputum for the preceding two weeks.

A rising creatinine, an active urine sediment that didn’t resolve with volume, and a systemic illness with hemoptysis is a pulmonary-renal syndrome until proven otherwise, and it moved malignancy and infection down the list. ANCA and PR3-ANCA came back positive, confirming granulomatosis with polyangiitis, GPA.

The eGFR was never the diagnostic point, it was the teaching point. The actual diagnostic work was being done by the urinalysis, read carefully and early, not by a calculated number that couldn’t have told you any of this.

Bob’s Pearls

1. An eGFR calculated from a rising creatinine isn’t a real number. The equations assume steady state; a moving creatinine hasn’t reached one.

2. A rising creatinine, by itself, doesn’t carry enough information to tell you how bad things are. Even complete, anephric kidney failure produces a rise from 0.8 to only about 1.5 to 2.0 over a day, since production is fixed once clearance hits zero. The same trajectory could reflect total kidney failure or something far milder, the number alone can’t distinguish them.

3. Don’t use the term “AKI” for a rising creatinine until you’ve considered volume and obstruction as possible causes. The label shouldn’t be applied before that reasoning is done, not after.

4. Volume contraction, obstruction, and intrinsic kidney injury can all produce the identical creatinine. A concentrated urine and a low-normal blood pressure make volume a reasonable hypothesis worth testing, even alongside other findings that also point elsewhere. Obstruction is quick to check, a bladder scan or renal ultrasound, but its pretest probability isn’t the same in everyone: much higher in older men with prostatic disease, lower, though not zero, in a woman in her 60s.

5. Any creatinine increase deserves a urinalysis. It’s cheap, fast, available at the bedside, and often more informative than the creatinine trend itself, protein, red cells, white cells, and casts tell you the kidney is inflamed, no equation does.

6. Dipstick protein is concentration-dependent, don’t take a 2+ at face value in a maximally concentrated urine. A specific gravity of 1.030 can inflate or mask a dipstick protein reading. An albumin to creatinine ratio corrects for concentration and gives you a real number.

7. When volume repletion doesn’t move the creatinine, believe the result. It’s not a failed intervention, it’s an important addition to the problem representation: whatever is happening, it isn’t volume alone, and the case now needs to be reframed accordingly.

8. Constitutional symptoms plus hemoptysis plus rising creatinine is a pulmonary-renal syndrome until proven otherwise. Malignancy and infection belong on that differential too, don’t let the eGFR distract from the story the patient is already telling you, and confirm with serology before naming the diagnosis.

This case was discussed at morning report and is presented as a true patient encounter, with the teaching points as they emerged in real time.

Daniel Kahneman and Gary Klein’s 2009 paper, “Conditions for Intuitive Expertise: A Failure to Disagree,” has had a real influence on how I think about clinical reasoning and the errors we all need to guard against [1]. What makes it different from most writing on the subject is that it refuses to pick a side. So much of the clinical reasoning literature is really a catalog of errors, heuristics and biases, cognitive traps to memorize and fear. That literature is valuable, and I have taught plenty of it myself. But a curriculum built only on what goes wrong teaches learners to distrust their own judgment without ever teaching them how good judgment actually gets built. This paper argues the other half deserves equal time: expert intuition is often genuinely, remarkably good, and it becomes good through a specific, learnable process. I want this piece, and the reasoning I teach going forward, to spend as much energy on that process as on the failure modes.

Two psychologists who expected to disagree

Daniel Kahneman spent his career in the “heuristics and biases” tradition, the school of thought that treats intuitive judgment with suspicion and goes looking for the mental shortcuts that lead it astray. He built that tradition together with Amos Tversky starting in the late 1960s, and it was their joint 1974 paper in Science that gave the field its name [6]. Their collaboration earned Kahneman the 2002 Nobel Prize in Economic Sciences for work he always described as fully joint. Tversky never formally shared it. He died in 1996, and Nobel rules do not allow posthumous awards. By the time Kahneman sat down to write this paper with Klein in 2009, he was in a real sense speaking for a field he had built with a partner no longer there to co-author it. Gary Klein came from “naturalistic decision making,” which studies fireground commanders, chess masters, and ICU nurses who make fast, high-stakes calls that turn out right far more often than chance would allow. When the two of them sat down to write this paper together, each expected to spend most of it arguing [1]. Instead they found they agreed on nearly everything, and the paper that resulted is one of the more useful things I have read on when a clinician’s gut feeling deserves trust.

Their shared starting point: intuitive judgment is recognition, not magic. An experienced fireground commander does not weigh options; he sees one plausible course of action, mentally tests it, and acts. A pediatric nurse who cannot explain why an infant looks wrong is still picking up on real, learnable cues, just ones she has never had to name out loud. The question worth asking is not whether recognition is happening. It always is. The question is whether the thing being recognized is real.

What has to be true for intuition to be trustworthy

Kahneman and Klein argue that skilled intuition depends on two conditions, and both have to hold [1]:

A high-validity environment. There has to be a stable, learnable relationship between the cues available and what actually happens next. Medicine and firefighting are offered as examples of environments with fairly high validity. Long-range political forecasting and picking individual stocks are examples of the opposite, because if a reliable public signal existed, the price would already reflect it.

Adequate opportunity to learn the regularities. Even in a high-validity environment, skill only develops with enough repetition and, critically, feedback that is fast and unambiguous. Chess masters need roughly a decade of serious play. A NICU nurse needs to have seen enough septic infants, and been told promptly when she was right or wrong, before the pattern becomes reliable. The relevant term here is deliberate practice, coined by Ericsson, Krampe, and Tesch-Römer to describe structured practice aimed at a specific weakness with fast, corrective feedback, as distinct from simply logging hours [2]. It is worth naming precisely because the two get conflated constantly. Volume alone, seeing a lot of patients without a mechanism for finding out which of your impressions were wrong, is not the same thing, and it is the distinction Kahneman and Klein lean on when they specify that opportunity to learn requires feedback that is rapid and unequivocal, not just repetition. It is also fair to say the deliberate practice literature has had a rockier decade than its early reputation suggests. Later meta-analyses found it explains a smaller share of expert performance than the original 1993 paper claimed, and a direct replication of that paper’s own violin study found a substantially weaker effect [2]. The core distinction, structured and corrected practice versus mere repetition, still holds up. The claim that it explains most of what separates experts from novices does not, and is worth holding loosely.

Medicine adds a wrinkle here worth naming honestly. A great deal of clinical feedback is delayed by days, weeks, or months, and that delay means most clinical experience is not deliberate practice in the strict sense above, since Ericsson’s definition depends on feedback fast enough to correct the very attempt that produced it. What delayed feedback is still good for, and why it does not fall into the same trap Kahneman and Klein warn about elsewhere, is worth its own section below.

Why confidence is not evidence

The finding I keep returning to is this: subjective confidence does not track accuracy. Kahneman and Klein show that confident, wrong judgments feel exactly like confident, right ones, because the feeling of certainty comes from how internally consistent the evidence seems, not from how good the evidence actually is [1]. A handful of weak findings that happen to point the same direction can generate as much conviction as several strong, independent ones.

They illustrate this with a story borrowed from Lewis Thomas, about an early 20th-century physician who developed a strong, repeatedly confirmed intuition for which ward patients were about to develop typhoid. He checked his hunch by palpating each patient’s tongue. Because he did not wash his hands between exams, he was spreading the disease himself, and his intuition kept being confirmed because he was making it true [1]. Kahneman and Klein call this a wicked environment, one where the feedback loop looks like it is teaching something real while it is actually being contaminated by the observer’s own actions. It is an uncomfortable story, and a useful one to sit with any time a clinician’s own behavior, who gets watched more closely, who gets the repeat test, who gets the benefit of the doubt, might be shaping the very outcome that seems to confirm the hunch.

Fractionated expertise

The idea in this paper I find most useful for teaching is what the authors call fractionated expertise. Reviewing the research on which professions actually develop reliable skill, they note that nurses, physicians, and auditors all appear on both the “experts do well here” list and the “experts do poorly here” list, depending on the specific task [1]. Expertise, in other words, is not a trait a person has or lacks. It is task-specific, and most professionals are genuinely skilled at parts of their job and not others, with no reliable internal signal telling them which is which.

That single idea reframes a lot of what gets called overconfidence in medicine. Consider a hospitalist who has read a thousand chest X-rays and reads them with real, earned skill, and then encounters a case of suspected IgG4-related disease they have seen perhaps twice in a career. The confidence they bring to the second case can feel identical to the confidence they bring to the first, because nothing internal marks the difference. It is not that the clinician is arrogant. It is that the confidence built by genuine, repetition-tested expertise in one domain does not know to stay in its lane when applied to another. The useful teaching point is not “distrust yourself more.” It is “learn to notice which kind of case you are in,” which is itself a skill that can be built and taught, not just a warning to be issued.

Problem representation and illness scripts: my road to expertise

Simon’s definition of skilled intuition, the one Kahneman and Klein build their whole argument on, is that a cue gives the expert access to information stored in memory, and that information supplies the answer. In clinical terms, the information stored in memory is the illness script: a structured mental package linking the conditions that predispose to a disease, the underlying fault, and the consequences that follow, refined case by case as a clinician accumulates experience [3]. But recognition cannot happen without something upstream of the script, which is a disciplined problem representation, the skill of translating a patient’s raw history and exam into an abstracted, precise summary rather than a running list of complaints [4]. “An elderly man with subacute, progressive dyspnea and orthopnea” is a representation that can be searched against stored scripts. A verbatim recitation of everything the patient said cannot.

This is where the delayed-feedback wrinkle from above actually gets resolved. A biopsy result that arrives after discharge cannot correct the exam performed that morning, so it fails the fast-feedback test deliberate practice requires. But Kahneman and Klein’s real objection to delayed feedback is not delay by itself. Their examples of untrustworthy intuition, long-range political forecasting and picking individual stocks, combine delay with ambiguity: the feedback, once it arrives, still does not clearly say whether the original judgment was right, because too many other variables intervened along the way. A late biopsy result does not usually have that problem. It is slow, but once it arrives it is typically unambiguous. That distinction, delay without ambiguity, is what makes clinical feedback still useful for a slower mechanism even where it cannot serve the fast one. It appears to be what builds and refines illness scripts over time, updating the stored link between a presentation and its actual outcome case by case. It also appears to be what teaches a clinician which features of a history or exam deserved more weight, shaping what gets folded into the next problem representation. This is my own extension of the framework rather than a claim Ericsson or Kahneman and Klein make directly, but it fits how encapsulation theory already describes clinical knowledge developing through incremental restructuring across cases rather than moment-to-moment correction [5].

Teach problem representation and illness scripts explicitly, rather than trusting them to develop by osmosis over a career, and you are teaching the actual machinery of expertise, closing the same feedback loop the paper describes, just on its slower of two tracks. This is why so much of my own teaching returns to these two concepts. They are the mechanism behind the recognition Kahneman and Klein describe in the abstract, made concrete enough to practice.

Teaching the skill, not just the failure

Kahneman and Klein spend as much energy explaining why expert intuition so often works as they do cataloging when it fails, and that balance is what I want to carry forward here. The failure modes are worth knowing. But if we only teach what goes wrong, we leave out the more useful half of the story: skilled clinical judgment is built, through real repetition in domains that actually reward it and through feedback, fast where it can be and delayed where it must be, that is honest enough to correct course. That is a positive curriculum, not just a defensive one, and it deserves at least equal billing.

If this whets your appetite, the clinical reasoning literature the paper itself points toward is a good place to go next. Jerome Groopman’s How Doctors Think and Atul Gawande’s Complications both wrestle with these same questions from inside the exam room, and Croskerry and Norman’s work on overconfidence in clinical decision making takes the argument further into medicine specifically than Kahneman and Klein do themselves [1].

Further reading

• Groopman J. How Doctors Think. Houghton Mifflin; 2007.

• Gawande A. Complications: A Surgeon’s Notes on an Imperfect Science. Profile Books; 2002.

• Croskerry P, Norman G. Overconfidence in clinical decision making. Am J Med. 2008;121(5 Suppl):S24-S29.

Full paper

Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. https://doi.org/10.1037/a0016755

References

[1] Kahneman D, Klein G. Conditions for intuitive expertise: a failure to disagree. Am Psychol. 2009;64(6):515-526. https://doi.org/10.1037/a0016755 [2] Ericsson KA, Krampe RT, Tesch-Römer C. The role of deliberate practice in the acquisition of expert performance. Psychol Rev. 1993;100(3):363-406. https://doi.org/10.1037/0033-295X.100.3.363 [3] Charlin B, Boshuizen HPA, Custers EJ, Feltovich PJ. Scripts and clinical reasoning. Med Educ. 2007;41(12):1178-1184. https://doi.org/10.1111/j.1365-2923.2007.02924.x [4] Bordage G, Lemieux M. Semantic structures and diagnostic thinking of experts and novices. Acad Med. 1991;66(9 Suppl):S70-S72. https://doi.org/10.1097/00001888-199109000-00045 [5] Boshuizen HPA, Schmidt HG. On the role of biomedical knowledge in clinical reasoning by experts, intermediates and novices. Cogn Sci. 1992;16(2):153-184. https://doi.org/10.1207/s15516709cog1602_1 [6] Tversky A, Kahneman D. Judgment under uncertainty: heuristics and biases. Science. 1974;185(4157):1124-1131. https://doi.org/10.1126/science.185.4157.1124

I’m writing this from Scotland, midway through a golf trip, having just walked off a links course with three other guys, all of us in our 70s, none of us moving like it. This is not a rare sighting here. St Andrews and the courses around it are full of septuagenarians and octogenarians who walk eighteen holes, carry or pull their own bags, and generally look — and seem to move through the world — younger than their age.

I’m 77. I’ve played golf for over 60 years. I notice this pattern every time I’m around a golf course full of older players, and every time, my first instinct is: golf keeps people young.

That instinct is also exactly what I’d expect to have if I were fooling myself. I love golf, I’ve built a good chunk of my identity and my social life around it for six decades, and I’ve spent enough time lately writing about confirmation bias in clinical guidelines to know how easily people — including me — find the evidence they’re already looking for. So before I say what I actually believe, I want to take the skeptical case seriously enough to try to knock my own thesis down.

The skeptical case, stated as strongly as I can

Here it is: golf courses aren’t a random sample of aging men. They select, hard, for people who can still walk, still swing a club without injury, and still afford the time and money for the game. A 77-year-old who’s frail, in pain, or managing a serious chronic illness isn’t out here. He’s already gone home, or he never took it up. So of course the population on the course looks younger than average — it would look that way even if golf itself did nothing at all for anyone’s health, purely because the unhealthy have already self-selected out. Epidemiologists call this healthy-user bias, and it shows up constantly in exercise research. The best golf-mortality data available — a Swedish registry study I’ll get to — cannot, by its own authors’ admission, rule this out. Neither can any of the other observational studies on the topic. None of it is a randomized trial. If someone wanted to dismiss the whole “golf is good for you” genre as survivorship bias wearing a lab coat, they’d have a real point.

I take that seriously. But I don’t think it’s the whole story, and here’s why.

Why I don’t think this is just selection

The best evidence here is Farahmand and colleagues’ study of the entire Swedish Golf Federation membership — over 300,000 golfers — matched against the national mortality registry. The overall finding was a 40% lower mortality rate than the general population, translating to roughly five additional years of life expectancy. That alone is consistent with pure selection. What isn’t as easily explained by selection alone is the internal structure of the result.

First, there’s a dose-response gradient: golfers with the lowest handicaps — the most skilled and, presumably, most active and frequent players — had the lowest mortality (SMR 0.53), while higher-handicap golfers had a smaller, though still real, benefit (SMR 0.68). Epidemiologists treat a graded dose-response relationship as one of the more persuasive signals for causation, precisely because simple selection (”healthy people play golf”) doesn’t obviously predict that more golf, or better golf, tracks with progressively lower mortality. Selection would have to operate not just at the gate — who takes up the game at all — but continuously and proportionally, sorting people by exactly how good and active they are. That’s not impossible, but it’s a more demanding version of the selection story than the simple one.

Second, the effect held up after adjusting for age, sex, and socioeconomic status, and it was consistent across every subgroup the authors checked. Consistency across strata is another of the classic markers researchers use to lean toward a real, generalizable effect rather than a quirk of one biased subpopulation.

Third, this study didn’t recruit self-selected enthusiastic responders the way a survey does — it’s a national registry matched to a national death registry, which removes a layer of self-selection bias that plagues a lot of exercise research.

None of this proves causation the way a trial would. But “40% lower mortality, graded by dose, robust across every subgroup, from national registry data” is a stronger pattern than pure survivorship bias predicts on its own, and I think it’s fair to weigh it that way.

What’s actually driving it

I don’t think golf has some unique magic. I think it’s a reasonably effective, low-injury-risk delivery mechanism for things we already have strong independent evidence matter: sustained low-to-moderate intensity movement, regular time outdoors, and — plausibly, though this part is more inferential — social contact with the same people on a recurring basis, which has its own well-documented mortality association independent of golf.

On the movement question, I want to be honest about my own game rather than the idealized one. Golfers vary a lot here — some always ride, some always walk, and I fall in between: I walk when the weather cooperates and ride when it’s hot, which in Alabama is a good part of the year. On the days I ride, it should, in theory, cut the exercise argument down substantially. It cuts it down less than I’d expect: even riding most of a round, I log a bit over two miles by the time I’ve walked cart-to-ball, around greens, and back, several times over eighteen holes. That’s not the four-to-six-mile walk of someone carrying a bag on foot — which is what I get on the days I walk — but it’s two-plus miles of intermittent movement even on my laziest days, multiple times a week, sustained into my late 70s. That’s itself a meaningful and fairly reliable predictor of health outcomes in the general exercise-mortality literature, separate from anything golf-specific. Even the “easy mode” version of the game still requires enough functional mobility that it’s delivering a real physiological dose, not just filtering for people who already have one.

Where I land

I think the honest reading of the evidence is that both things are happening, but they’re not equal contributors. Selection is real — I’m not naive about who ends up still playing golf at 77 — but the size, gradient, and consistency of the mortality effect in the best available data go beyond what selection alone would predict. I think golf is doing some of the work itself, mostly by making sustained low-to-moderate activity and outdoor time easy to keep doing for decades, in a way that a lot of other exercise regimens don’t survive into a person’s 80s.

That’s a real conclusion, not a hedge, and I want to hold myself to it: if someone showed me golfers had identical mortality to a matched group of similarly mobile, similarly wealthy non-golfers doing no other structured activity, I’d need to revise this. Nobody has run that comparison. Until they do, I’m going with the data I have — and, for what it’s worth, with sixty years of experience suggesting the same thing.

Sources

• Farahmand, B., Broman, G., De Faire, U., Vågerö, D., & Ahlbom, A. (2009). Golf: a game of life and death – reduced mortality in Swedish golf players. Scandinavian Journal of Medicine & Science in Sports, 19(3), 419–424.

• American Heart Association (2020). Want to live longer? Get into the swing of golfing — coverage of Dr. Adnan Qureshi’s Cardiovascular Health Study analysis, presented at the International Stroke Conference.

• Holt-Lunstad, J., Smith, T. B., Baker, M., Harris, T., & Stephenson, D. (2015). Loneliness and Social Isolation as Risk Factors for Mortality: A Meta-Analytic Review. Perspectives on Psychological Science, 10(2), 227–237.

Bottom Line

Legionella survives inside the very immune cells sent to kill it, which is why a patient with severe pneumonia who isn’t improving on a beta-lactam should make you think of it. The organism uses a molecular syringe to hijack the alveolar macrophage that swallows it, turning a phagocyte into a private incubator. That single mechanistic fact explains the clinical picture, the antibiotic choice, and why New York City just announced a new round of cooling tower regulations after a summer outbreak on the Upper East Side.

As of the end of August, that cluster had produced 94 cases and 11 deaths, and the city’s investigators are now trying to match the genetic strain from patient samples to the cooling towers in the neighborhood. It is a useful moment to walk through what actually happens in the lung, because the diagnostic and treatment logic for Legionella all flows from one idea: this bacterium doesn’t live in the space between your cells. It lives inside them.

The patients behind this post

Three patients over the past two years made this teaching stick in a way the textbook description never did, precisely because only one of them was easy. The first presented with prominent diarrhea alongside an obvious pneumonia, and that combination made the diagnosis quick. The other two were trickier. In the second, diarrhea came first and was confusing on its own, with the fever and dyspnea, and the pneumonia on chest x-ray, only showing up later that same evening. The third was the hardest: high fevers up to nearly 105°F with a normal chest x-ray. It took a CT scan to reveal the pneumonia, and only after that did the diarrhea start. He also had elevated transaminases, and mild hyponatremia showed up in all three. Legionella coverage was started empirically before any confirmatory testing came back on the third patient, not because of a specific clue pointing to it, but because he looked sicker than the working picture explained.

The big point isn’t any single finding. It’s that Legionella presents in enough different ways that it deserves a place in the differential earlier than most of us reach for it. Three patterns, in particular, should move it up the list sooner rather than later: a pneumonia that doesn’t look like typical community-acquired pneumonia, diarrhea accompanied by significant fever even without an obvious pulmonary source yet, and high fevers that are simply confusing, not fitting a clean clinical story. None of these is specific to Legionella on its own. What they share is that they’re all reasons to widen the differential rather than wait for a more classic presentation to declare itself.

Why beta-lactams fail

Most bacterial pneumonias are fought with cell-wall-active drugs like penicillins and cephalosporins, which work because the bacteria are sitting in the extracellular space where the drug can reach them. Legionella doesn’t cooperate. Once inhaled in an aerosol and phagocytosed by an alveolar macrophage, it deploys a type IV secretion system, essentially a needle that injects more than 300 bacterial proteins into the host cell, that stops the phagosome from fusing with the lysosome. Instead of being digested, the bacterium remodels its compartment into a Legionella-containing vacuole, a protected niche where it replicates undisturbed.

A drug that can’t get inside the macrophage can’t touch the organism, no matter how susceptible it looks on a plate. This is why beta-lactams, despite often showing in vitro activity, fail clinically, and why treatment requires antibiotics that concentrate intracellularly: respiratory fluoroquinolones (levofloxacin) or macrolides (azithromycin). Guidelines for severe community-acquired pneumonia generally favor a respiratory fluoroquinolone or a macrolide, sometimes in combination, once Legionella is suspected or confirmed. Doxycycline is a reasonable alternative when those first-line agents are contraindicated, poorly tolerated, or interact with a patient’s other medications; it shares the same advantage of good intracellular penetration. The evidence behind it is thinner, mostly case series and in vitro data suggesting somewhat lower bactericidal activity than the fluoroquinolones, so it’s a second-line option rather than a co-equal one.

The clinical picture that should raise suspicion, and why imaging can mislead

Legionella has historically been grouped under “atypical pneumonia,” a label that undersells how sick these patients get. What actually points toward it is a pattern of findings outside the lung layered on top of a pneumonia that isn’t behaving, and the frequencies are worth knowing rather than treating this as a vague gestalt: high fever with rigors and drenching sweats in up to 90% of cases, confusion, gastrointestinal symptoms (nausea or watery diarrhea) in roughly 30-50%, which can precede the respiratory complaints entirely, elevated transaminases commonly enough that they’re a real part of the picture (the literature is consistent that this is frequent, though it doesn’t pin down a precise percentage the way it does for fever and GI symptoms), and hyponatremia. None of these is individually sensitive or specific enough to diagnose the disease, but the combination, especially in a patient with a smoking history, an immunocompromising condition, a history of malignancy, or a relevant water exposure (a hotel stay, a recent hospitalization, a cruise), should prompt specific testing rather than empiric broad-spectrum coverage alone.

The mechanism of the hyponatremia is still debated. The traditional explanation is inappropriate ADH secretion, but some more recent work points toward a renal salt-wasting process instead. Either way, it’s a clue that doesn’t need to be dramatic to count.

Pneumonia itself is not always an easy diagnosis, and it’s worth being honest about what each tool actually does. Chest x-ray functions less as a screening test and more as a confirming test: it’s good at confirming what you already suspect, but a clean film doesn’t mean the lung is clean, especially early in the course or in a patient whose findings are subtle. CT is where confusing cases get sorted out, which is exactly what happened in the third patient above, though that’s a general pneumonia finding rather than something specific to Legionella. Point-of-care ultrasound can also help, particularly when a consolidation sits against the pleural surface where the probe can see it, but it’s operator-dependent and won’t catch everything a CT would. None of these is a substitute for the others; the point is to keep going when the clinical picture doesn’t match a negative film, not to assume any single tool has ruled the diagnosis out.

One imaging fact that is Legionella-specific: radiographic resolution lags clinical improvement by 5 to 7 days even with effective treatment. A patient who’s clinically better shouldn’t be expected to have a clean film to match; the chest x-ray takes longer to catch up, sometimes months to fully clear. Don’t let a slow-to-resolve film undermine confidence in a treatment that’s working.

Tests to Consider

• Urine antigen test: Fast, cheap, and the most commonly used test, but it has a real blind spot. Pooled sensitivity is roughly 80% overall and closer to 85-90% specifically for Legionella pneumophila serogroup 1, with specificity near 100%. The catch is that it only reliably detects serogroup 1, which causes the large majority of culture-confirmed cases but not all of them. A negative urine antigen does not rule out Legionella caused by other serogroups or non-pneumophila species.

• Sputum culture on selective media (BCYE agar): The traditional gold standard, and the only test that yields an isolate for outbreak investigations, which is exactly what’s happening in New York right now as they try to match patient strains to specific cooling towers. Its practical limitation is that a substantial fraction of patients, especially those with a dry cough, can’t produce an adequate lower respiratory sample.

• PCR of respiratory specimens: Highly sensitive and specific, and unlike the urine antigen test, it isn’t limited to serogroup 1. Increasingly used where available, including on multiplex respiratory panels.

• A pragmatic bedside filter: One French cohort found that a CRP below roughly 130 mg/L made a positive urine antigen essentially very unlikely, with a strong negative predictive value in that population. It’s a single-center finding that needs replication before it becomes a rule, but it’s a reasonable sanity check: a pneumonia with a low CRP is less likely to be Legionella, all else being equal.

Best practice, when Legionella is genuinely on the differential, is to send urine antigen and a respiratory specimen for culture or PCR simultaneously rather than relying on any single test.

Why cooling towers, and why now

Legionella is a freshwater organism that becomes dangerous when it finds a warm, stagnant, aerosol-generating environment to amplify in, and a cooling tower (the large evaporative structure that dissipates heat from a building’s air conditioning system) is close to ideal. New York’s new measures target the parts of that lifecycle that are actually fixable: closing the gap between the Health Department’s cooling tower registry and the Department of Buildings’ permitting system so towers can’t quietly go unregistered, accelerating enforcement hearings, and exploring faster culture methods so a cluster can be confirmed and traced sooner. The city’s own numbers make the enforcement gap concrete: as of late August, only 81% of registered cooling towers citywide were in compliance with monthly testing requirements. The city is also convening a scientific panel to look at alternative cooling technologies, an acknowledgment that as heat waves get longer, the conditions that favor Legionella growth are becoming more common, not less.

Cooling towers are the classic exposure, and the reason they drive outbreak investigations, but they’re far from the only route. Anything that generates a fine water aerosol from a warm, stagnant reservoir is a candidate: hotel and building hot water systems, whirlpool spas and hot tubs, decorative fountains, and humidifiers all show up repeatedly in case reports and outbreak investigations. Travel is a recognized risk category in its own right, not because travel itself does anything, but because hotels, resorts, and cruise ships concentrate exactly the kind of complex water infrastructure, hot tubs, large plumbing systems, sometimes their own cooling towers, that Legionella favors, and travelers often can’t tell you anything more specific than where they stayed. A hotel stay in the Middle East is exactly the kind of history that should prompt asking about hot tubs, spa use, or unusual water exposure during the trip, even when the patient can’t point to anything specific.

A brief boat ride is a different story, and it’s worth naming directly: the exposure in that case was never pinned down. Since salinity normally suppresses Legionella’s growth, plain ocean spray isn’t the expected route; a heated seawater spa or a vessel’s technical water system (non-potable seawater used for deck washdown or engine cooling, which the CDC flags as its own aerosolization risk on ships) are plausible mechanisms, but there was no way to confirm which, if either, applied here. That’s the more honest and more common experience than the tidy exposure-to-diagnosis stories usually told: not every patient has a classical, confirmable exposure, and the diagnosis shouldn’t wait on finding one. A classic exposure history is comforting to have when it’s present, but its absence is not a reason to exclude Legionella from the differential.

One exposure worth being precise about: standard window air conditioning units don’t pose a Legionella risk, because they cool with refrigerant rather than water and don’t generate an aerosol. The exception is an evaporative cooler (”swamp cooler”), which does use a water reservoir and can harbor Legionella if poorly maintained; these are sometimes described loosely as window units, which can create confusion in a travel or exposure history. When a patient’s only reported exposure is “a window AC unit,” it’s worth asking specifically whether it was actually an evaporative system, and if not, whether there’s another source, a nearby cooling tower, a hot tub, a humidifier, that better fits what we know about how the organism actually spreads.

None of that changes how you manage the patient in front of you, but it’s the backdrop worth knowing when a cluster like this makes the news and patients start asking questions.

One more number worth carrying into the differential: Legionella is likely missed in roughly half of actual cases, and some estimates put the miss rate considerably higher. A Connecticut surveillance study found only about 1 in 10 statistically predicted cases were ever clinically diagnosed; an older Ohio study found correct diagnosis in only about 3% of sporadic cases. The gap exists because testing has to be specifically requested and specifically thought of, not because the disease is truly rare. Every pearl above is really in service of that one fact: the diagnosis is missed far more often from not looking than from the organism being hard to catch once you do.

Bob’s Pearls

1. Legionella survives inside alveolar macrophages by blocking phagosome-lysosome fusion. That single fact explains why beta-lactams don’t work and why macrolides or respiratory fluoroquinolones do.

2. Legionella presents in enough different ways that it deserves an earlier spot in the differential, not a later one. Consider it sooner for: pneumonia that doesn’t look like typical CAP, diarrhea with significant fever, or high fevers that are simply confusing.

3. High fever with rigors in up to 90% of cases, GI symptoms in 30-50% that can precede the pneumonia entirely, and commonly elevated transaminases. None of these alone is diagnostic, but the combination should raise suspicion.

4. Pneumonia isn’t always an easy diagnosis. Chest x-ray confirms more than it screens; CT sorts out the confusing cases; POCUS can help but won’t catch everything. Once treatment is working, don’t expect the film to catch up quickly, radiographic resolution lags clinical improvement by 5 to 7 days.

5. A classic exposure history is comforting to have when it’s present. Its absence is not a reason to exclude Legionella from the differential.

6. Legionella is likely missed in roughly half of actual cases, possibly more. The gap is mostly a failure to test, not a hard diagnosis to make once you look.

Further reading

• CDC, Laboratory Testing for Legionella: https://www.cdc.gov/legionella/php/laboratories/index.html

• StatPearls, Legionnaires’ Disease: https://www.ncbi.nlm.nih.gov/books/NBK430807/

• Rello J et al., “Severe Legionnaires’ disease,” Annals of Intensive Care, 2024. PMC10987467

• CDC MMWR, cruise ship whirlpool spa outbreak investigation (Legionella, 1994): https://www.cdc.gov/mmwr/preview/mmwrhtml/00032265.htm

• Tomic Linsak D et al., “Sea water whirlpool spa as a source of Legionella infection,” Journal of Water and Health, 2021.

• CDC, Control Legionella guidance for cruise ship operators (technical water aerosolization risk): https://www.cdc.gov/control-legionella/php/guidance/cruise-ship-operators.html

• Whitney CG et al., “Estimating the True Burden of Legionnaires’ Disease,” Connecticut surveillance study, 2019. PMID 31225857

• NYC Health Department / Mayor’s Office press release, September 1, 2026, on new Legionella prevention measures following the Upper East Side cluster.

If you caught my first post on primary aldosteronism — whether you listened or read it — you may remember the central takeaway: this disease is far more common than most of us were taught, hypokalemia is a late and unreliable sign, and the aldosterone-to-renin ratio (ARR) is the recommended screening test. All of that was grounded in a 2018 podcast with Dr. David Calhoun.

Today’s episode takes things further — and into more unsettled territory. This is a 2020 Annals on Call episode built around a study called “The Unrecognized Prevalence of Primary Aldosteronism: A Cross-Sectional Study,” published in the July 7, 2020 issue of Annals of Internal Medicine. My guest was Dr. Jennifer Brown, who at the time was a clinical and research fellow in cardiovascular medicine at Brigham and Women’s Hospital in Boston and was the lead author of the study.

I want to be transparent with you about something before we go further. Some of what Dr. Brown presents — particularly around the limitations of the ARR and a different way of thinking about screening — is intellectually compelling and directionally likely to be correct, but is not yet fully validated clinical guidance. I find it persuasive. I also want you to be able to distinguish between what is established and what is an emerging research framework, so you can apply your own judgment. That distinction will become important as we go.

The Problem: A Gap Between Prevalence and Diagnosis

Dr. Brown opened by naming the motivation behind her study. Primary aldosteronism (PA) is associated with substantial cardiovascular risk — risk that goes beyond what you would expect from the elevated blood pressure alone. People with PA have roughly two to four times higher odds of stroke, atrial fibrillation, coronary artery disease, and heart failure compared to patients with ordinary primary hypertension matched for blood pressure. Effective treatments exist. So if you can identify these patients, you can do something meaningful for them.

The trouble is that we are not identifying them. At the time of this study, evidence suggested that even among patients with resistant hypertension — the group most clearly recommended for screening — only 2 to 3 percent of eligible patients were actually being tested.

Why? A few reasons that probably feel familiar. Many clinicians still anchor on the classic picture of PA: severe hypertension, profound hypokalemia, and a unilateral adenoma visible on imaging. If that’s your mental model, you’re only going to test when you see those features — and you’ll miss most cases. There’s also a perception that testing is technically onerous. And then there’s the question of the screening test itself — which is where the study makes its most provocative contribution.

The Study: Pooling Five Populations Across the Blood Pressure Spectrum

To understand what Dr. Brown’s group found, you need to understand what they did — and the important limitation it carries.

They pooled five separate study protocols. Four were volunteer studies investigating hormonal mechanisms of blood pressure regulation. The fifth was a protocol embedded in the UAB Resistant Hypertension Clinic. Together, these gave them populations spanning the full blood pressure range, from normotensive individuals through stage 1 and stage 2 hypertension to resistant hypertension.

The key measurement in each protocol was 24-hour urine aldosterone after oral sodium loading. Why this approach? When you load someone with sodium, you suppress the renin-angiotensin-aldosterone axis. Renin goes down, angiotensin II goes down, and aldosterone should follow. Any aldosterone measured in that context — when renin is already suppressed — is by definition renin-independent. That is the fundamental biochemical abnormality of primary aldosteronism. The 24-hour urine measurement also provides a more stable picture than a single blood draw, which is highly variable.

Here is the limitation to hold in mind: these were largely research volunteer populations, not representative samples from a primary care clinic. Self-selected research participants are systematically different from your average patient with newly diagnosed stage 1 hypertension. The study is best understood as hypothesis-generating — a careful look at the physiology of aldosterone across the blood pressure spectrum — rather than as a prevalence study that tells us exactly how common PA is in a general practice.

The Continuum, and What It Suggests

What they found was that rather than a clean separation between people who had PA and people who did not, there was a continuum. Across the entire blood pressure spectrum, renin-independent aldosterone production ranged from barely detectable all the way to well above current diagnostic thresholds — and the distribution shifted as blood pressure severity increased.

In other words: PA does not appear to be a categorical disease that is either present or absent. It looks more like a spectrum — from subclinical or mild excess autonomous aldosterone production, to moderate, to overt. And that spectrum tracks with blood pressure severity.

This maps onto analogies we already use. Heart failure is not heart failure — we recognize NYHA class I through IV. Someone with mild exertional limitation is not the same as someone who is symptomatic at rest, but they exist on the same pathophysiologic continuum. Primary aldosteronism may work similarly. The person with resistant hypertension, severe hypokalemia, and a large adenoma is at one end. There may be a person with stage 1 hypertension, a normal potassium, and mildly autonomous aldosterone production at the other end — who never gets diagnosed because we are only looking for the dramatic presentation.

What is the biology? The most compelling hypothesis involves aldosterone-producing cell clusters (APCCs) — small islands of cells within the adrenal cortex, identifiable on immunohistochemistry, that express aldosterone synthase at higher-than-normal levels and harbor somatic mutations. They may produce aldosterone in a partially autonomous manner, explaining why the phenomenon could be so common without requiring discrete adenomas in most patients. This is still speculative; we do not yet know how much clinical impact these clusters have in vivo. But it is an appealing and biologically plausible mechanism.

The ARR: A Useful but Imperfect Test

Here is where the conversation gets important — and where I want to be careful about how I frame things for you.

Dr. Brown’s group asked: how well does the ARR perform as a screening test? They used the 24-hour urine aldosterone after sodium loading as their reference standard. Against that reference, they found that the ARR had a sensitivity of only 25 to 50 percent. In resistant hypertension specifically, the negative predictive value was only 80 percent — meaning that if you test a resistant hypertension patient with the ARR and it comes back negative, you may still be wrong one in five times.

This is a striking finding, and it should make us less confident in a negative ARR than most of us currently are. But I want to be honest about what it does and does not show. The study measured how often the ARR agreed with one particular physiologic test — 24-hour urine aldosterone after sodium loading. That test is a reasonable and physiologically sound approach to measuring renin-independent aldosterone production. It is not, however, the clinical gold standard for PA diagnosis, which has traditionally involved a combination of confirmatory suppression testing and adrenal vein sampling. A different choice of reference standard might produce different sensitivity figures.

What the Brown study does establish — convincingly — is that a negative ARR does not reliably exclude PA, particularly in high-risk patients. The test is useful, and you should still use it. But a negative result should reduce, not eliminate, your suspicion.

A Different Way of Thinking: Renin Suppression as the Signal

Dr. Brown’s key reframing was to move attention away from the ratio and toward renin suppression as the primary signal. The logic is physiologically grounded: the defining abnormality of PA is aldosterone produced independently of renin. So the fundamental question to answer is whether renin is suppressed.

If renin is not suppressed, aldosterone is probably responding appropriately to the renin-angiotensin system. Lower suspicion for PA.

If renin is suppressed, something is driving aldosterone production in the absence of the normal stimulus. That is the abnormal physiology you are looking for.

She proposed a practical three-zone framework based on the aldosterone level in the context of suppressed renin:

Aldosterone above 15 ng/dL with suppressed renin: High suspicion for overt PA. Refer to endocrinology. Consider imaging and adrenal vein sampling. Surgery is a possibility if a unilateral source is identified.

Aldosterone between 5 and 15 ng/dL with suppressed renin: A gray zone — possible mild to moderate renin-independent aldosterone excess. Empiric mineralocorticoid receptor antagonist therapy may be reasonable; subspecialty referral is another option.

Aldosterone below 5 ng/dL with suppressed renin: Lower suspicion for clinically significant PA, though suppressed renin still warrants attention.

I find this framework intuitive and physiologically appealing. I should also tell you what it is not: it has not been validated in prospective outcomes trials. We do not yet have a study showing that applying this framework and treating the gray zone empirically produces better cardiovascular outcomes than treating patients with a standard antihypertensive to similar blood pressure targets. That evidence does not yet exist.

There is also an important caveat about suppressed renin that this framework requires you to hold: approximately 25 to 30 percent of people with essential hypertension have low-renin hypertension, particularly Black patients and older adults, without any autonomous aldosterone production. Suppressed renin alone does not equal PA. It is the starting point for a more careful look — not a diagnosis.

Practical Testing: Simpler Than You Think

One of the most actionable points from Dr. Brown was that the testing barrier is largely a myth we have constructed for ourselves.

You do not need to take patients off their antihypertensive medications before measuring aldosterone and renin. Washing off medications is difficult, dangerous, and often impossible in the patients who need testing most. You do not need to worry about the time of day or special patient preparation.

Sit the patient down in the phlebotomy chair. Draw a plasma renin and a plasma aldosterone. That’s it.

A single blood measurement of aldosterone is imperfect and variable. But it is far better than not testing at all because the protocol seemed too complicated. If you find a suppressed renin with a non-trivial aldosterone, you have started the conversation. You can repeat the measurement. You can refer to endocrinology. The important thing is that you identified the signal.

For inpatient physicians: this is testable in the hospital. It may not be confirmatory, but drawing renin and aldosterone during an admission for a patient with severe or resistant hypertension is a reasonable and useful starting point that plants the flag for the outpatient team.

Why We Are Slow to Use MRAs

I asked Dr. Brown a question that reflects a real tension in how we practice. We know that patients with resistant hypertension often go on spironolactone empirically and do well. Shouldn’t we be doing this more broadly, and earlier, even in stage 1 or stage 2 hypertension where we suspect mild PA?

Her answer was candid. One barrier is that most patients with mild hypertension respond adequately to any first-line agent — a calcium channel blocker, an ACE inhibitor, a thiazide. Their blood pressure comes down, and there is no apparent reason to do additional testing or change the drug. The issue, she argued, is that controlling blood pressure with the wrong drug is not the same as treating the underlying disease. If someone has autonomous aldosterone excess and you put them on amlodipine, their blood pressure may improve — but the aldosterone is still stimulating mineralocorticoid receptors throughout the cardiovascular system, promoting fibrosis and adverse remodeling. You have treated the number without addressing the pathophysiology.

The other common barrier is fear of hyperkalemia. Physicians who remember the reports of dangerous hyperkalemia after spironolactone was more widely adopted in heart failure patients are understandably cautious. For patients with overt PA — where autonomous aldosterone is driving significant renal potassium wasting — the hyperkalemia risk is genuinely lower than in heart failure, and Dr. Brown makes this point well. That said, if the continuum model is correct and we extend treatment to patients with mild autonomous aldosterone excess who may not have significant potassium wasting, that reassurance does not transfer as cleanly. Monitoring potassium remains appropriate across the board, and the risk profile in mild or subclinical PA is not yet well characterized.

Where We Stand in 2025

This episode aired in 2020, and the field has continued to move — though perhaps not quite as far as some headlines suggest.

Three major guideline bodies — the 2025 Endocrine Society, the 2024 European Society of Cardiology, and the 2025 AHA/ACC — have all opened the door to broader screening. The Endocrine Society issued a conditional recommendation to consider screening all adults with hypertension, regardless of severity. The ESC gave a Class IIa recommendation to consider screening all adults with confirmed hypertension. The AHA/ACC gave a Class IIb recommendation for considering screening in stage 2 hypertension, on top of a Class I recommendation for resistant hypertension.

It is worth being precise about what this convergence means. Conditional recommendations and Class IIb designations are the guideline panels’ way of saying: the evidence points in this direction, but it is not yet strong enough to recommend universally. These are cautious, hedged signals from experts who reviewed the same literature — not a unified endorsement of universal screening. The direction is clear; the certainty is not.

Despite even the existing recommendations, fewer than 2 percent of eligible patients are currently being screened. Whatever your threshold for acting on the new guidance, there is enormous room to close the gap with what we already know.

One additional update: a 2025 study in Annals of Internal Medicine by Leung et al., paired with an editorial by Cohen, found that the saline infusion suppression test — the most commonly used confirmatory test — performed poorly in distinguishing PA from essential hypertension. There is a growing view that in patients with a compelling clinical picture, formal confirmatory testing may not always be necessary before initiating MRA therapy. This is consistent with the renin-centric approach Dr. Brown advocated — but the same caveat applies: this is an emerging position, not yet standard of care.

Bob’s Pearls

1. Primary aldosteronism looks more like a spectrum than a switch. The Brown study, while conducted in research populations that may not generalize directly to your clinic, makes a compelling case that renin-independent aldosterone production exists along a continuum that tracks with blood pressure severity. A patient without hypokalemia and without dramatic hypertension can still have meaningful autonomous aldosterone production.

2. A negative ARR does not exclude PA, particularly in high-risk patients. In Dr. Brown’s analysis, the ARR had a sensitivity of only 25 to 50 percent against a physiologic reference standard. The exact number depends on the populations and reference test used, but the takeaway is clear: a negative ARR should lower your suspicion, not close the case.

3. Suppressed renin is the key signal — but not sufficient on its own. Low-renin hypertension is common, particularly in Black patients and older adults, without PA. Suppressed renin tells you to look more carefully at the aldosterone level. It is not a diagnosis.

4. Testing is simpler than we make it. No medication holds. No special timing. Draw renin and aldosterone in a seated patient. Imperfect data, collected and acted on, is far better than no data because the protocol seemed too hard.

5. The renin-centric three-zone framework is physiologically appealing — and not yet validated by outcomes data. Think of it as a useful way to structure your reasoning about the aldosterone and renin levels you collect, not as a proven clinical algorithm. The direction it points is likely correct. The evidence to fully support it is still being built.

This podcast aired in July 2020. The article is: Brown JM et al., “The Unrecognized Prevalence of Primary Aldosteronism: A Cross-Sectional Study,” Annals of Internal Medicine, 2020. Updates reflect the 2025 Endocrine Society Clinical Practice Guideline on Primary Aldosteronism, the 2024 ESC guideline, the 2025 AHA/ACC hypertension guideline, and Leung et al., Annals of Internal Medicine, 2025.