What is a differential diagnosis? Method and examples

A differential diagnosis is the list of conditions that could plausibly explain a patient's presentation — and the reasoning process by which that list is generated, ranked and narrowed to a working diagnosis. It is built from the history and examination, then pruned with targeted investigation. The abbreviation is DDx. Confusingly, clinicians also use the term for a single item on the list ("costochondritis is a differential here"), so context matters. The discipline is not the list itself but the habit of holding several explanations open until the evidence closes them.

This is a reference for qualified clinicians and students: what a differential diagnosis is, why it matters for diagnostic safety, the frameworks used to generate one, how to narrow it rationally, the reasoning errors that derail it, and how to document it. It is educational material, not patient-facing advice, and it does not replace your local or national diagnostic pathways.

What is a differential diagnosis?

A differential diagnosis is the set of conditions that could account for a patient's symptoms and signs, together with the systematic process of distinguishing between them. Because most symptoms are non-specific and few single tests are definitive, clinicians work from a list rather than a hunch — ordering it by probability and by danger, then testing to collapse it.

The textbook sequence is well described by the US National Library of Medicine's patient-facing explainer, which breaks the process into four steps: gather clues from the history, perform a physical examination, assemble the list of possible conditions, then order tests to confirm or rule them out — testing urgent possibilities first (MedlinePlus). Its worked example is instructive precisely because it is so ordinary: fatigue, it notes, can be caused by anaemia, depression, heart disease, thyroid disease or sleep disorders, among many other conditions. One symptom, many mechanisms — which is the whole reason the method exists.

Two terminological points are worth fixing. First, a differential diagnosis is not the same as a working or provisional diagnosis: the working diagnosis is the single explanation you are currently treating, while the differential is everything still on the table. Second, DDx is the conventional abbreviation; you will also see ddx, DD and D/Dx in notes and teaching material.

Why does the differential diagnosis matter?

Because diagnosis fails often enough to be a leading patient-safety problem, and a narrow differential is one of the mechanisms by which it fails. The World Health Organization puts diagnostic errors at 5–20% of physician–patient encounters and reports that most people will experience a diagnostic error in their lifetime. A disciplined differential is the main cognitive defence against that.

• Scale. The WHO patient safety fact sheet (11 September 2023) states that diagnostic errors occur in 5–20% of physician–patient encounters, that doctor reviews have found harmful diagnostic errors in at least 0.7% of adult hospital admissions, and that most people will experience a diagnostic error in their lifetime.
• Where it happens. The same WHO fact sheet notes that some estimates suggest as many as 4 in 10 patients are harmed in primary and ambulatory settings, and that up to 80% of that harm may be avoidable — which locates much of the diagnostic risk in exactly the outpatient consultation where differentials are built fastest and with least data.
• What counts as an error. The US National Academies of Sciences, Engineering, and Medicine defined diagnostic error from the patient's perspective: the failure to establish an accurate and timely explanation of the patient's health problem, or to communicate that explanation to the patient. It also concluded that most people will experience at least one diagnostic error in their lifetime (Improving Diagnosis in Health Care, 2015, ch. 3).
• The hidden denominator. Autopsy studies have found that about 9% of patients had a major diagnostic error that went undetected during life, according to the AHRQ PSNet primer on diagnostic errors — a reminder that clinicians rarely learn about their own misses, and so get very little corrective feedback.

Note the definitional nuance in the National Academies framing: timeliness and communication are part of diagnostic accuracy. A differential that eventually reaches the right answer three months late, or that is never explained to the patient, still counts as a failure.

How do you build a differential diagnosis?

Generate broadly, then narrow deliberately. In practice: define the problem precisely, generate candidates with a systematic framework, rank them by probability and by danger, look for discriminating features on history and examination, then test only where the result would change what you do.

• 1. Define the problem. Convert the complaint into a precise problem representation — not "chest pain" but "sudden tearing interscapular pain in a 58-year-old with uncontrolled hypertension". Age, sex, duration, tempo, and context carry most of the diagnostic information, and a sharper problem statement produces a sharper list.
• 2. Generate candidates systematically. Use a framework (below) rather than free recall, which is biased toward whatever you saw or read most recently.
• 3. Rank on two axes. Order by probability (what is common in this population) and separately by danger (what would be catastrophic to miss). These are different lists, and both must be addressed.
• 4. Seek discriminating features. Target history and examination at the findings that actually separate your top candidates, rather than re-collecting data that fits them all equally.
• 5. Test to change management. Investigate where a result shifts probability across a decision threshold. Low-yield reflex panels add cost and false positives without moving the diagnosis.
• 6. Keep the list open. Record what remains possible, define what would make you revisit, and arrange explicit follow-up — safety-netting is part of the differential, not an afterthought.

What frameworks help you generate the list?

Three families of framework dominate teaching, and experienced clinicians switch between them. All of them exist for the same reason: to make the list wider than unaided recall would make it, before you start cutting it down.

• Anatomical / by system. Work outward from the anatomy of the complaint. For chest pain: cardiac, vascular, pulmonary, gastrointestinal, musculoskeletal, psychological. This is the framework used in our chest pain differential diagnosis reference, and it is hard to beat for regional symptoms.
• Pathological category (the "surgical sieve"). Run the problem past each class of disease mechanism — vascular, infective, inflammatory, neoplastic, degenerative, traumatic, metabolic/endocrine, autoimmune, congenital, iatrogenic and drug-related, functional. Taught in many mnemonic forms (VINDICATE and VITAMINS being the common ones), its value is the prompt, not the acronym: it catches the category you would otherwise skip — most often the iatrogenic one.
• The three-question triad. Ask: what is most likely, what is most dangerous, and what is most treatable? The second question is the one that drives safe practice, because a low-probability, high-lethality diagnosis (aortic dissection, meningitis, ectopic pregnancy) still earns active exclusion.
• Illness scripts. Expert diagnosis leans heavily on stored patterns — typical patient, typical time course, typical findings — matched against the case. Scripts are fast and usually right, which is both their strength and their risk: they work by resemblance, and resemblance can mislead.

A caution on mnemonics: they are scaffolding for an unfamiliar or stalled problem, not a ritual to perform on every sore throat. When you already hold a good condition-specific differential, use that instead — the generic sieve is for when your list looks implausibly short.

How do you narrow the differential?

By changing probabilities, not by collecting facts. Each finding and each test moves a candidate up or down; your job is to identify the findings with real discriminating power, apply them, and stop when one explanation is probable enough — and the dangerous alternatives unlikely enough — to act on.

• Start from a pre-test probability. Prevalence in your population and setting is the anchor. The same presentation carries a very different differential in a paediatric clinic and a geriatric ward.
• Prefer findings with real discriminating power. Some features shift probability substantially; most do not. Where published likelihood ratios exist for a presentation, they tell you which questions are worth asking — in chest pain, for example, pain radiating to both arms carries a positive likelihood ratio of roughly 7 for acute myocardial infarction.
• Test thresholds, not curiosity. Order a test when the result crosses a management threshold in either direction. Testing without a specific indication is low-yield; in fatigue, the AAFP review notes that unindicated laboratory testing may change treatment in only 5% of patients.
• Exclude the dangerous before the plausible. A benign explanation that fits is not a reason to stop looking for the lethal one that also fits.
• Re-examine when the course deviates. Failure to improve as predicted is diagnostic information. Non-response is the commonest signal that the working diagnosis is wrong.

A worked example: fatigue (illustrative)

Fatigue is a useful teaching case because it is both extremely common — among the top 10 reasons patients attend primary care, per the American Family Physician review — and almost entirely non-specific. The following sketch is illustrative only, written to show the method rather than to guide the care of any real patient.

Take a 34-year-old with four months of tiredness, no weight loss, no fever, and normal observations. A structured approach puts the candidates into tiers rather than a flat list. The AAFP review classifies fatigue as physiologic, secondary (to another condition) or chronic, and assesses physiologic causes first (Latimer, Gunther & Kopec, 2023).

• Most likely (physiologic and common secondary): insufficient or fragmented sleep, overwork and deconditioning, depression or anxiety, iron-deficiency anaemia, hypothyroidism, medication effect (including over-the-counter and herbal agents), early pregnancy, post-viral and post-COVID states.
• Most dangerous (low probability, high consequence): occult malignancy, significant anaemia from chronic blood loss, heart failure, diabetes with metabolic decompensation, adrenal insufficiency, obstructive sleep apnoea with cardiovascular consequences.
• Most treatable (worth excluding early because the fix is simple): iron deficiency, thyroid dysfunction, vitamin B12 deficiency, a culprit drug, untreated depression, sleep apnoea.
• Discriminating questions: snoring and witnessed apnoeas; menstrual blood loss; mood, anhedonia and sleep architecture; postexertional malaise (central to ME/CFS); red-flag features — weight loss, night sweats, lymphadenopathy, bleeding.

What the method buys you here is restraint in both directions: it prevents a reflex broad panel on a well patient, and it prevents the opposite error of writing off four months of fatigue as stress without asking the handful of questions that would separate the tiers.

What reasoning errors derail a differential?

Most diagnostic failures are not knowledge gaps about rare disease; they are failures of process on common presentations. The AHRQ PSNet primer on diagnostic errors groups the classic cognitive failures into four patterns, all of which act by truncating the differential too early.

• Anchoring and premature closure — holding to the initial impression despite contradictory information, and stopping the differential before it is complete. This is the single most important pattern to guard against.
• Availability heuristic — judging the current patient by recent, memorable cases rather than by base rates.
• Framing effects — allowing how the case was handed over ("another anxious frequent attender") to shape the list before you have seen the patient.
• Blind obedience — over-relying on a test result or a senior opinion, for instance accepting a false-negative rapid test over a convincing clinical picture.

The system side matters just as much. The same primer identifies weak teamwork and communication, and unreliable processes for test-result follow-up and telephone triage, as major contributors — and notes that clinicians often never learn of their own diagnostic errors, so the feedback loop that would correct them barely exists.

One widely repeated claim deserves qualifying. It is often said that diagnostic error is a failure of fast, intuitive "System 1" thinking. A 2024 review in the Journal of Evaluation in Clinical Practice argues otherwise: errors originate from both System 1 and System 2 thinking, and typically occur because of a lack of access to the appropriate knowledge (Norman et al., 2024). The practical implication is unflattering to bias-spotting workshops: reorganising and strengthening knowledge is more useful than trying to think slowly on command.

How should you document a differential diagnosis?

In the Assessment, with your reasoning visible. A defensible entry states the working diagnosis, names the serious alternatives you considered and why you ruled them in or out, records the uncertainty honestly, and specifies the safety-net: what should prompt the patient to return, and what you will review.

In SOAP terms, the differential belongs in A, bridging the data in Subjective and Objective and the actions in Plan. Compare "Chest pain", which documents nothing, with an entry that shows the reasoning (illustrative, and written to the pathway in use — your thresholds and tests are the ones that apply):

Two points about that entry matter more than its wording. First, each dangerous alternative is closed by a test or finding, not by an impression — a "PE unlikely" risk score is a reason to proceed to D-dimer, never a rule-out on its own, and the absence of hypoxia excludes nothing. Second, the uncertainty and the safety-net are written down, because a differential that stays open needs a documented plan for revisiting it. Our guide to how to write a SOAP note covers the structure in detail, the free printable SOAP note template gives you a layout to work from, and SOAP vs POMR compares note formats if you are choosing between them.

Where does AI decision support fit?

At the generation step, where human recall is weakest. A tool that proposes candidates you had not considered and flags the dangerous ones addresses the specific mechanism — a list that was too short — behind a large share of diagnostic error. It does not address the part only a clinician can do: weighing this patient's story, examining them, accepting accountability for the decision.

Shifaa AI — an AI clinical assistant for doctors, currently in private beta — includes differential diagnosis support that surfaces a ranked differential with the red-flag possibilities raised first and the guidance behind each suggestion shown, so you can check it rather than trust it. It is decision support, not diagnosis: the clinician reviews, edits and signs. Honest framing on the limits is worth reading before relying on any such tool — see can AI diagnose patients? and, for the evidence picture on computerised diagnostic support, note that the AHRQ primer still describes it as unproven for improving overall diagnostic accuracy.

Frequently asked questions

What is the difference between a differential diagnosis and a diagnosis?

A differential diagnosis is the list of conditions still capable of explaining the presentation, held open while you gather evidence. A diagnosis is the single explanation you have concluded is correct and are acting on. The working diagnosis sits between them: your current best explanation, with the differential still formally open.

How many conditions should be on a differential diagnosis?

There is no fixed number. In practice three to five serious candidates, plus any high-lethality condition that must be actively excluded, is a workable working list. Quality matters more than length: a short list that includes the dangerous possibilities beats a long list padded with implausible ones.

What does DDx stand for?

DDx is the standard abbreviation for differential diagnosis, from "D" for diagnosis and the delta-like "Dx" shorthand. You will also see ddx, DD and D/Dx. Clinicians sometimes use "a differential" for a single candidate on the list as well as for the list itself.

Can AI generate a differential diagnosis?

AI tools can generate and rank candidate diagnoses from a clinical summary, which is genuinely useful at the point where human recall narrows too soon. They cannot take the history, examine the patient or carry the accountability. The AHRQ PSNet primer still describes computerised diagnostic support as unproven for improving overall accuracy, so treat output as a prompt to think, never as a verdict.

References

• MedlinePlus (US National Library of Medicine). What is a differential diagnosis? medlineplus.gov
• World Health Organization. Patient safety fact sheet, 11 September 2023. who.int
• National Academies of Sciences, Engineering, and Medicine. Improving Diagnosis in Health Care (2015), ch. 3 — Overview of Diagnostic Error in Health Care. nationalacademies.org
• Agency for Healthcare Research and Quality, PSNet. Diagnostic Errors primer (last reviewed 2024). psnet.ahrq.gov
• Norman G, Pelaccia T, Wyer P, et al. Dual process models of clinical reasoning: the central role of knowledge in diagnostic expertise. J Eval Clin Pract. 2024;30(5):788–796. AHRQ PSNet summary
• Latimer KM, Gunther A, Kopec M. Fatigue in Adults: Evaluation and Management. Am Fam Physician. 2023;108(1):58–69. aafp.org
• Outpatient evaluation of chest pain / diagnosis of acute coronary syndrome. American Family Physician. aafp.org