A doctor walks into a room, sits down and asks, "So what brings you in today?" Then she listens for a minute or two without interrupting, which is rarer than it should be. By the time she reaches for the stethoscope, she has already narrowed a hundred possible explanations to a handful. The tests she orders afterwards are not a fishing trip; each is meant to move a probability.
That is diagnosis: not a flash of genius but a running estimate that rises and falls as evidence arrives, made by a human mind with a fast pattern-matcher and a slow checker, both of which err. All cases in this piece are invented, and every number is an illustration. This explains how medicine works and is not medical advice, and it is not a tool for diagnosing yourself or anyone else. For a health question, see a clinician; in an emergency, call your local emergency number. If you or someone you know is in a mental-health crisis, findahelpline.com lists services by country.
A classic study by John Hampton and colleagues (British Medical Journal, 1975) followed 80 patients sent to a medical outpatient clinic. After the referral letter and the history, the clinician's diagnosis matched the final one in 66 of 80 cases, about 82%; the examination added little, and tests mostly confirmed. A later replication in Salt Lake City (Peterson and colleagues, 1992) found about 76%. These are small, old studies of a particular clinic, and they are the source of the saying that "listen to the patient, he is telling you the diagnosis", so treat the exact percentages as a rough indicator, not a law.
The reason is simple. A story contains a lot of information: how fast it started, what makes it better or worse, what the person was doing, what else runs in the family. The examination tests specific ideas the history raised. Tests then confirm or refute them, and each should be chosen because its result would change what you do. A useful habit has a name: hypothetico-deductive reasoning. Form a short list of explanations early (often within the first minute), then ask what finding would separate them.
Before a test, the clinician has a pre-test probability: how likely this patient is to have the condition, given age, history, setting and the local base rate. A test then shifts it. The best summary of a test's power is its likelihood ratio (LR): how much more common the result is in people with the condition than in people without it. It comes from the two familiar numbers.
The picture was put on a single chart by Fagan in 1975: a nomogram where a straight line from the pre-test probability through the LR gives the post-test probability. You can do the same with a slider.
An invented test for an unnamed condition. "Positive twice" assumes two independent tests, which real tests rarely are. Illustration, not a calculator for any real test or person.
Take a pre-test probability of 10% and a test with 90% sensitivity and 95% specificity. The LR+ is 18 and the LR− is 0.105. A positive result moves 10% to about 67%; a negative one moves it to about 1.2%; two independent positives would take it to 97%. Notice what a positive did not do: it did not make the answer certain. And with a pre-test probability of 1%, the same positive only reaches about 15%, which is the screening lesson of Epidemiology (35/58) in another form.
A reference range is usually set to include the middle 95% of healthy people. That means 5% of perfectly healthy people fall outside it on any one test, by definition. Order several tests and the chance that at least one is flagged climbs.
Independent tests in a healthy person. Real results are partly correlated, which lowers the chance somewhat, but not to zero. Illustrative.
With 20 independent tests and 95% ranges, a healthy person has about a 64% chance of at least one flag; with 40, about 87%. Widen each range to cover 99% of healthy people and 20 tests flag at least one in only 18% of cases. This is why clinicians do not order "everything" for someone who feels well: the abnormal results are more often noise than disease, and each one invites a follow-up with its own harms. A test is ordered when the result would change what happens next.
The differential diagnosis is the ranked list of explanations still in play. Two proverbs frame it. "When you hear hoofbeats, think horses, not zebras" (attributed to a Maryland professor, Theodore Woodward, in the 1940s) says to start with what is common. Hickam's dictum answers that a patient can have as many diseases as they please. A third rule overrides both: always ask what you must not miss. A rare diagnosis that can kill within hours and is easy to treat earns a place on the list even at a low probability.
That last rule is a number. In 1980 Stephen Pauker and Jerome Kassirer described the threshold approach: you should treat when the probability of disease passes the point where expected benefit to those who have it equals expected harm to those who do not.
If the treatment is gentle and the disease dangerous, the threshold is low and you treat on a hunch. If the treatment is risky and the disease mild, the threshold is high and you want more certainty. Between a lower and an upper probability a test is worth doing, because a positive could push you over the threshold or a negative could pull you under it.
Harm and benefit are in the same invented units; the test is an invented one with LR+ 10 and LR− 0.1. A teaching picture of the Pauker and Kassirer idea, not a decision tool.
With harm 5 and benefit 45, the treatment threshold is 10%. A test that moves odds ten-fold up or down is worth doing when the probability is between about 1.1% and 52.6%. Below it, even a positive result would stay under the threshold. Above it, even a negative result would stay over it. Change the harm to 20 and the benefit to 20: the threshold becomes 50% and the testing zone runs from about 9% to 91%. The bigger the harm of treatment relative to the benefit, the more evidence you need first.
Experienced clinicians often diagnose in seconds: the "look" of the patient or the pattern of the story matches an illness script built from hundreds of cases. This is fast, usually right and hard to explain, which psychologists call System 1 (see The Biases That Fool Everyone, 17/02, for the general idea). The other mode is slow and analytic: list the possibilities, test them against the findings, think about what would make them wrong. Researchers such as Pat Croskerry have argued that most diagnostic errors involve a failure of the fast mode that the slow mode did not catch; others, notably Geoff Norman and colleagues, point out that experts also err when they are slow and that knowledge gaps matter. The field has not settled which "fix" works best.
The named failures are well known. Anchoring: sticking with the first idea. Premature closure: stopping once a plausible answer appears. Confirmation bias: giving more weight to findings that fit. Availability: overestimating what you saw last week. Base-rate neglect: forgetting how common the condition is. Debiasing, such as the habit of asking "what else could this be?" or "what is the worst thing it could be?", has some support but mixed evidence in trials. Try an invented case with fictional diseases.
Invented patient "Ms Ardent". Three fictional possibilities: common, mild Brannock fever (prior 90%), serious but rare Tolliver's disease (5%), and something else (5%). Reveal findings one at a time.
All diseases, findings and numbers are invented to show the process. The "anchored" bar discounts findings that favour the other disease by the weight you set. Not a model of any real clinician.
Weighed fully, the findings (fever, mild cough, swollen joints, then a strongly positive Test Z) take Tolliver's disease from 5% to 74%. A clinician who gives only 30% weight to findings against Brannock fever ends the same case at about 9% Tolliver, 90% Brannock: the story is still "a common mild illness". The error did not need a wrong fact; it needed only a discount on the awkward ones. A simple defence is to say out loud, "what finding would make me change my mind?" and then check whether it has appeared.
Because diagnosis is a process, error is hard to count: it needs a gold standard such as an autopsy or follow-up. The US National Academy of Medicine's 2015 report Improving Diagnosis in Health Care defined diagnostic error as the failure to establish an accurate and timely explanation of the patient's health problem, or to communicate it to the patient, and concluded that most people will experience one at least once. A review by Berner and Graber (2008) put erroneous physician diagnoses at about 10 to 15%. A 2023 modelling study led by David Newman-Toker at Johns Hopkins (published in BMJ Quality & Safety) estimated that about 795,000 Americans a year die or are permanently disabled by diagnostic error: 371,000 deaths and 424,000 disabilities, with vascular events, infections and cancers accounting for three-quarters. That is an estimate with wide uncertainty from one country, built on models, and earlier estimates ranged from 40,000 to 4 million, so treat it as an order of magnitude and not a count.
The causes are mixed: faulty thinking, but also missing information, busy rooms, hand-offs, tests not followed up, and patients who were not heard. Error is as much a matter of systems as of individuals, which is why patient safety work looks at how a hospital runs, not only at who was tired.
Computers can now match specialists on some narrow tasks. In 2018 the US FDA authorised the first AI system to diagnose a specific eye disease, diabetic retinopathy, from retinal photographs. Systems read mammograms, skin photographs and scans, though a 2020 claim that an algorithm could beat expert radiologists drew replies from researchers who said the paper lacked the detail needed to reproduce it. For general reasoning, large language models can score well on case vignettes. In a 2024 randomised trial in JAMA Network Open (Goh and colleagues), physicians given access to an LLM scored 76% on diagnostic reasoning against 74% for those using usual resources, a difference that was not statistically significant, while the model alone scored higher than both groups. A good tool on its own is not the same as a good tool in a doctor's hands.
The limits are specific. A model trained at one hospital can fail at another with different scanners, patients or prevalence, which is called dataset shift. Few studies reporting AI accuracy have been tested outside the place they were built. A widely used commercial sepsis-warning score, when tested at one US university hospital system (Wong and colleagues, JAMA Internal Medicine, 2021), had an area under the curve of 0.63, missed 67% of patients who developed sepsis and still raised an alert on 18% of hospitalisations. And people tend to over-trust a confident machine (automation bias). The base-rate arithmetic from earlier still applies:
An invented AI with 92% sensitivity and 90% specificity where it was built (prevalence 20%). Numbers per 1,000 people flagged or not. Illustration, not any real product.
Where it was built, at 20% prevalence, 70% of its alerts are real (184 true, 80 false per 1,000). Used at 1% prevalence with the same accuracy, only 8.5% of alerts are real (about 9 true and 99 false per 1,000). Lose 10 points of specificity in the new setting and it falls to 4.4%. The software did not get worse; the world around it changed. Most clinical AI is therefore best judged by prospective tests in the setting where it will be used, and by whether patients do better.
Diagnosis is a probability that moves with each piece of evidence, steered by a mind that is fast, usually right and predictably biased. A test is worth doing when its result could change what happens next. The best defences are a short list, a worry about what you must not miss, and the question "what would change my mind?"
Try this: next time you make a confident guess about anything (why the car will not start, why a friend is late) write down your first idea and two others, and the one finding that would make you drop the first. Then see whether you look for it.
Sources. Likelihood ratios, Bayes in diagnosis and the threshold idea (Fagan 1975; Pauker and Kassirer 1980) are standard clinical-decision-making teaching. History-taking: Hampton et al., BMJ 1975 (66 of 80); Peterson et al., West J Med 1992 (about 76%), via secondary summaries. Diagnostic-error definition: National Academy of Medicine (2015); Berner and Graber (2008), via Wikipedia, Diagnostic error. Newman-Toker et al., BMJ Qual Saf 2024;33:109–120 (online July 2023). Goh et al., JAMA Netw Open 2024;7(10):e2440969. Wong et al., JAMA Intern Med 2021 (AUC 0.63; 67% of sepsis missed; alerts on 18%), via PubMed. FDA 2018 retinopathy authorisation: Wikipedia, Artificial intelligence in healthcare. All cases are invented; all interactive numbers were computed in node. This is not medical advice.