Article type: explainer
Scope: General education about Bayesian reasoning, diagnostic uncertainty, cognitive bias, and patient-clinician communication; not individualized medical advice.
Last updated: August 3, 2026
SEO title: Bayes’ Theorem, Rare Diagnoses, and Medical Error
Medical students are often taught some version of the same rule:
When you hear hoofbeats, think horses, not zebras.
The lesson is sensible. A cough is more likely to be caused by a respiratory infection than a rare autoimmune disorder. A headache is more likely to be a migraine or tension headache than a brain tumor. A tired patient is more likely to be sleep-deprived, stressed or anemic than to have a condition documented only a few thousand times.
Doctors cannot investigate every imaginable disease every time someone walks into an urgent care center. If they did, medicine would become slower, more expensive and potentially more dangerous. Unnecessary testing produces false positives, incidental findings, radiation exposure, invasive follow-up procedures and enormous anxiety.
The problem is not that doctors begin with the common explanation.
The problem begins when they stop thinking after finding one.
“Common things are common” is a starting probability—not a final diagnosis.
Doctors Are Usually Right to Start With the Common
A physician examining a patient is working under uncertainty. The patient’s true condition is initially unknown, so the doctor begins with what statisticians call a prior probability: the estimated likelihood of a diagnosis before all the case-specific evidence has been considered.
Age, medical history, local disease patterns, recent exposures and the frequency of particular illnesses all influence that starting estimate.
This is not lazy medicine. It is rational medicine.
Doctors also use heuristics—mental shortcuts developed through training and repeated clinical experience. These shortcuts allow clinicians to recognize familiar patterns quickly. The Agency for Healthcare Research and Quality notes that physicians frequently use heuristics to form provisional diagnoses, especially when patients arrive with common symptoms. (PSNet)
Without pattern recognition, even routine medical care would become unmanageable.
But a heuristic is supposed to accelerate thinking, not replace it.
When a Useful Shortcut Becomes an Intellectual Trap
The same mental processes that allow an experienced doctor to identify influenza, appendicitis or a urinary tract infection quickly can also produce diagnostic errors.
One of the most important is anchoring: becoming attached to the first plausible diagnosis and failing to adjust when new information points elsewhere.
A related problem is premature closure. The physician finds an explanation that appears adequate, accepts it and effectively closes the investigation before the diagnosis has been sufficiently tested. AHRQ describes this as relying on an initial impression despite later information that contradicts it. (PSNet)
There is also the representativeness error: expecting every disease to present in its textbook form. A doctor may recognize a heart attack when the patient has crushing chest pain but hesitate when the same condition presents through nausea, fatigue, shortness of breath or unusual discomfort.
These errors do not necessarily arise because a physician is unintelligent, careless or malicious. They often arise because human beings naturally prefer a coherent answer over unresolved uncertainty.
Once a familiar diagnosis has been attached to the patient, later evidence can be unconsciously interpreted in ways that preserve it.
The fever is “just part of the infection.”
The unexplained weakness is “probably anxiety.”
The abnormal test is “likely incidental.”
The patient’s failure to improve means the medication “needs more time.”
Each explanation may be reasonable by itself. Collectively, however, they can become a structure built to protect a diagnosis that is no longer adequately explaining the case.
Bayes’ Theorem: Start With Probability, Then Update It
Bayes’ theorem is often presented as an intimidating mathematical formula. Its central idea is much simpler:
Begin with what was likely before seeing the evidence, and then revise that probability according to how well the new evidence supports or contradicts it.
In clinical medicine, this generally involves three components:
- The prior probability: How likely was the disease before considering the new finding?
- The strength of the evidence: How much more likely would this symptom, test result or pattern be if the patient had the disease?
- The posterior probability: How likely is the disease after incorporating that evidence?
Medical researchers commonly express this relationship as:
Post-test odds = pretest odds × likelihood ratio
The likelihood ratio represents how strongly a finding changes the probability of a diagnosis. A highly distinctive symptom or test result can dramatically increase the probability of an initially uncommon condition. Conversely, evidence that should be present but is absent can reduce it. (PubMed Central (PMC))
Imagine that a particular condition has only a 1 percent probability at the beginning of an evaluation. That is low, and a doctor would be justified in considering more common explanations first.
Now imagine that the patient has a finding associated with a positive likelihood ratio of 20. The estimated probability rises from 1 percent to roughly 17 percent.
A second finding with a likelihood ratio of 10 could—assuming it provides substantially independent information—raise the probability to approximately 67 percent.
The diagnosis did not become more common in the general population. It became more likely in this particular patient.
That distinction is the heart of Bayesian medicine.
Population rarity tells a doctor where to begin. Patient-specific evidence determines where the reasoning should go next.
“Rare” Does Not Mean “Statistically Irrelevant”
Doctors sometimes respond to a patient’s concern by saying that the suspected condition is extremely rare.
That may be factually true. It may also be logically incomplete.
The relevant question is not simply:
How rare is this condition in the general population?
It is:
How rare is this condition among people with this patient’s specific combination of symptoms, test results, history and response to treatment?
A condition affecting one person in 10,000 is unlikely in a randomly selected person. It may be far more likely among patients with a distinctive clinical pattern.
There is also a scale problem hidden inside the word rare. The FDA reports that more than 10,000 recognized rare diseases collectively affect over 30 million Americans—approximately one in ten people in the United States. Each disorder may be unusual individually, but rare disease as a category is not rare at all. (U.S. Food and Drug Administration)
And rare diseases are only part of the problem. Common dangerous diseases can themselves present in uncommon ways.
The “zebra” may sometimes be a genuinely rare disorder. At other times, it is an ordinary disease wearing an unusual disguise.
Every Uncommon Case Has to Appear Somewhere
A disease that occurs once in every 20,000 patients still occurs.
When it does, the patient does not arrive carrying a sign that says rare case. They arrive with fatigue, dizziness, abdominal pain, fever, weakness, vomiting, numbness or some other symptom shared by dozens of more common conditions.
And whoever evaluates that patient may be encountering that particular presentation for the first time.
That is the statistical reality physicians must respect:
If an uncommon event occurs, it will necessarily occur in front of someone who is accustomed to seeing the common event instead.
The doctor’s lack of prior personal experience does not reduce the probability created by the patient’s actual evidence.
“Never seen it before” is information about the clinician’s experience. It is not proof about the patient’s condition.
The Detail That Does Not Fit May Be the Most Important Detail
Good diagnosis is not merely the search for an explanation that fits most of the symptoms. It is also the search for evidence that the leading explanation cannot comfortably explain.
A common diagnosis may account for the fever and cough but not the neurological change.
It may explain the stomach pain but not the falling blood pressure.
It may explain fatigue but not progressive muscle weakness.
It may explain an initial symptom but not its persistence, recurrence or severity.
In Bayesian terms, these deviations are new evidence. They must change the probability distribution.
The same is true when treatment fails. If a diagnosis predicts that a patient should improve after a particular intervention and the patient does not improve, that failure is not merely an inconvenience. It is diagnostic information.
It may not immediately prove that the diagnosis is wrong. Treatments fail for many reasons. But it should reduce confidence in the original hypothesis and increase consideration of alternatives.
A diagnosis that can survive every contradiction is not a medical hypothesis. It is a belief.
Emergency Medicine Shows Why Presentation Matters
A major AHRQ systematic review estimated that approximately 5.7 percent of emergency department visits involve at least one diagnostic error—roughly one in 18 visits if the estimate is generalizable to the United States. The review identified stroke, heart attack, aortic aneurysm or dissection, spinal cord injury and blood clots among the conditions most vulnerable to serious misdiagnosis-related harm. (PSNet)
The review also demonstrated how strongly presentation affects recognition. The average estimated missed-diagnosis rate for stroke was 17 percent, but it was approximately 4 percent when patients presented with weakness and 40 percent when they presented with dizziness or vertigo. The same disease became much harder to recognize when it did not resemble the mental prototype clinicians expected. (NCBI)
These estimates have important methodological limitations, which the report itself acknowledges. They should not be interpreted as proof that every emergency department operates at the same error rate. They do, however, expose a central weakness in diagnostic reasoning: physicians are much better at recognizing diseases when those diseases behave as expected.
A separate study of 2,428 hospitalized adults who died or required transfer to intensive care found diagnostic errors in 23 percent of the reviewed cases. Because the study examined an unusually sick population, that figure cannot be applied to all hospitalized patients. It nevertheless found that problems involving clinical assessment, test selection and test interpretation were major areas for improvement. (JAMA Network)
Listening to the Patient Is Part of the Evidence
Patients are not automatically correct because they have researched their symptoms. Online self-diagnosis can exaggerate rare possibilities, ignore base rates and cause people to interpret vague symptoms as evidence of frightening diseases.
Bayes’ theorem cuts both ways. It warns doctors not to dismiss strong evidence merely because a disease is uncommon. It also warns patients that a rare disease remains unlikely when the evidence supporting it is weak or nonspecific.
But there is an important difference between a patient declaring, “I know I have this rare disease,” and a patient saying:
- “This symptom was present before the medication.”
- “The pain is in a different location than the one you described.”
- “I have completed the treatment, but the symptoms are becoming worse.”
- “This does not feel like the previous episodes I have had.”
- “There is a part of my history that I do not think has been accounted for.”
Those statements are not competing medical degrees. They are data.
The National Academies has emphasized that diagnosis is a collaborative process and that patients and families are central members of the diagnostic team. Its landmark report concluded that most people are likely to experience at least one meaningful diagnostic error during their lifetime. ([National Academies][7])
Listening to patients does not require surrendering clinical judgment. It means recognizing that the patient has continuous access to information the physician sees only briefly.
What Better Diagnostic Reasoning Looks Like
The solution is not to order every available test or place every rare disorder on every differential diagnosis. That would create another form of bad medicine.
The solution is structured uncertainty.
A responsible clinician can say:
“This is the most likely diagnosis based on what we know now, but it is provisional.”
Before closing the case, the doctor can ask:
- What finding does not fit the leading diagnosis?
- What dangerous alternative would be costly to miss?
- What evidence would cause me to reconsider?
- Has the patient’s condition followed the expected course?
- Am I discounting the patient’s concern because of how it was expressed rather than because of its medical content?
Patients can ask equally reasonable questions:
- “What parts of my presentation support this diagnosis?”
- “Is there anything that does not fit it?”
- “What should happen if the treatment does not work?”
- “Which symptoms should cause me to return immediately?”
- “At what point would another diagnosis or specialist referral become appropriate?”
These questions do not demand that the doctor chase zebras. They keep the stable door open when the hoofbeats stop sounding like horses.
Bayes’ Theorem Is Humility Expressed as Mathematics
The deepest lesson of Bayes’ theorem is not a formula. It is an ethic of intellectual revision.
A good doctor does not abandon base rates. Common diseases should usually remain at the top of the initial differential diagnosis.
But a good doctor also does not worship base rates. The prior probability must yield to accumulating evidence.
Medicine fails when “most likely” quietly becomes “certain,” when contradictory details are treated as annoyances, or when a patient’s persistence is mistaken for irrationality.
The uncommon should not be presumed.
It should not be hunted indiscriminately.
But neither should it be dismissed merely because the person examining it has never personally seen it before.
Common things are common.
Rare things still happen.
And when they happen, they do not arrive with the word rare stamped across the patient’s forehead.
They first arrive looking almost ordinary—except for the details that someone decided not to ignore.
—
Frequently Asked Questions
How is Bayes’ theorem used in medical diagnosis?
Bayes’ theorem helps clinicians revise the estimated probability of a disease as new information becomes available. A doctor begins with a pretest probability based on prevalence and patient characteristics, then updates it using symptoms, examination findings and test results.
Does Bayes’ theorem mean doctors should test for every rare disease?
No. Testing people with extremely low pretest probabilities can generate false positives and unnecessary procedures. Bayesian reasoning supports selective testing when the patient’s specific evidence meaningfully raises the probability of an uncommon diagnosis.
What is anchoring bias in medicine?
Anchoring occurs when a clinician becomes attached to an initial diagnosis and does not adequately revise it after receiving contradictory information. It is closely related to premature closure, in which the diagnostic process ends before the leading explanation has been sufficiently verified.
Can a rare disease become the most likely diagnosis?
Yes. A disease can be rare in the general population but comparatively likely in a patient with a highly distinctive combination of symptoms, risk factors, test results or treatment failures.
—
References and Further Reading
Diagnostic Error and Patient Safety
- Newman-Toker, D.E., Peterson, S.M., Badihian, S., et al. “Diagnostic Errors in the Emergency Department: A Systematic Review.” Agency for Healthcare Research and Quality, Comparative Effectiveness Review No. 258, December 2022.
A comprehensive review of diagnostic-error frequency, commonly missed conditions, atypical presentations and resulting patient harms in emergency departments.
- National Academies of Sciences, Engineering, and Medicine. “Improving Diagnosis in Health Care.” National Academies Press, 2015.
A landmark report examining the diagnostic process, patient participation, clinical reasoning, health-system failures and strategies for reducing diagnostic error.
- Agency for Healthcare Research and Quality Patient Safety Network. “Diagnostic Errors.” Last reviewed June 2024.
An accessible overview of diagnostic errors, anchoring, availability bias, premature closure and system-level contributors to missed or delayed diagnoses.
- Auerbach, A.D., Lee, T.M., Hubbard, C.C., et al. “Diagnostic Errors in Hospitalized Adults Who Died or Were Transferred to Intensive Care.” JAMA Internal Medicine, vol. 184, no. 2, 2024, pp. 164–173.
A multicenter study examining diagnostic errors and related harm among hospitalized adults who died or required intensive care.
Bayesian Reasoning in Medicine
- Webb, M.P.K., and Sidebotham, D. “Bayes’ Formula: A Powerful but Counterintuitive Tool for Medical Decision-Making.” BJA Education, vol. 20, no. 6, 2020, pp. 208–213.
A clear explanation of prior probability, likelihood, posterior probability and the interpretation of medical tests.
- Nixon, M.P., Momotaz, F., Smith, C., et al. “From Pre-Test and Post-Test Probabilities to Medical Decision Making.” BMC Medical Informatics and Decision Making, vol. 24, article 210, 2024.
Explores how Bayesian probability can be connected to real clinical decisions involving the benefits and costs of testing and treatment.
- National Academies of Sciences, Engineering, and Medicine. “The Diagnostic Process.” In Improving Diagnosis in Health Care, 2015.
Reviews diagnostic hypothesis formation, verification, probabilistic reasoning and the need to revise diagnoses as information accumulates.
Cognitive Bias and Clinical Judgment
- Saposnik, G., Redelmeier, D., Ruff, C.C., and Tobler, P.N. “Cognitive Biases Associated With Medical Decisions: A Systematic Review.” BMC Medical Informatics and Decision Making, vol. 16, article 138, 2016.
Reviews evidence concerning anchoring, overconfidence, availability bias and other cognitive influences on physician decision-making.
- Agency for Healthcare Research and Quality Patient Safety Network. “From Possible to Probable to Sure to Wrong—Premature Closure and Anchoring in a Complicated Case.”
A clinical case analysis demonstrating how a plausible early diagnosis can persist despite accumulating contradictory evidence.
Rare Diseases
- U.S. Food and Drug Administration. “Rare Diseases at FDA.”
Provides current federal definitions and estimates concerning the number of rare diseases and the population collectively affected by them.
- NIH Genetic and Rare Diseases Information Center. “Genetic and Rare Diseases Information Center.”
A federal resource for patients, families and clinicians seeking reliable information about recognized rare conditions, diagnostic resources and specialist organizations.
—
Medical disclaimer: This article is intended for general education and discussion. It does not diagnose any condition or replace evaluation by a qualified health professional.
