| name | design-differential-diagnosis |
| description | Use when generating and systematically ranking the possible diagnoses for a patient's presenting symptoms and clinical findings |
| source | Harrison's "Principles of Internal Medicine" (21st ed.); Kassirer & Kopelman "Learning Clinical Reasoning" (1991); Sackett "Evidence-Based Medicine" (2000) clinical decision-making |
| tags | ["medicine","diagnosis","clinical-reasoning","differential-diagnosis"] |
| verified | true |
Design Differential Diagnosis
Generate a comprehensive, prioritized differential diagnosis using systematic anatomical, physiological, and probabilistic frameworks, then narrow it with targeted evaluation.
Why This Is Best Practice
Adopted by: USMLE clinical reasoning assessment, Royal College of Physicians and Surgeons certification (Canada, UK), Johns Hopkins/Mayo Clinic clinical reasoning curricula, WHO IMCI (Integrated Management of Childhood Illness) diagnostic algorithms.
Impact: Systematic differential diagnosis reduces diagnostic error — which affects 12 million Americans/year (IOM 2015) — by 40% compared to pattern-recognition-only approaches (Graber et al. Arch Int Med 2005); premature closure (settling on a diagnosis too early) accounts for 36% of diagnostic errors.
Why best: Structured differential generation forces clinicians to consider diagnoses outside their initial pattern match, preventing anchoring bias and premature closure — the two most common cognitive errors in diagnosis. Probabilistic ranking ensures the most dangerous conditions are evaluated first.
Sources: Harrison's 21st ed. Part 1; Kassirer & Kopelman (1991) ch. 3–5; Sackett et al. (2000) ch. 3; Graber et al. Arch Intern Med 165:1493–1499 (2005).
Steps
-
Extract the problem representation — create a one-sentence clinical summary: "A [age][sex] with [key risk factors] presents with [duration] [chief complaint] plus [2–3 key associated findings], most notable for [pivotal finding]." This activates the correct disease schema.
-
Generate the initial differential broadly — list all conditions that could explain the chief complaint; use a systematic framework to avoid omission:
- Anatomical: work through each organ system that could produce this symptom
- Pathophysiological (VITAMIN-D): Vascular, Infectious, Traumatic, Autoimmune/Metabolic, Iatrogenic/Idiopathic, Neoplastic, Degenerative/Drug, Congenital/Endocrine
- Pattern-based: serious conditions not to miss + most common conditions for this demographic.
-
Apply pre-test probability — for each candidate diagnosis, estimate base rate given: age, sex, risk factors, geographic prevalence, and referral context. High prior probability diagnoses deserve more prominent position even with fewer specific findings.
-
Identify pivot features — find 2–3 findings that most powerfully discriminate between diagnoses: features that are highly specific (LR+>10) for one diagnosis or highly sensitive (LR-<0.1) for ruling out another. Pivot features drive the diagnostic workup.
-
Rank the differential in three tiers:
- Must-not-miss: life-threatening or limb-threatening conditions with even low probability — evaluate first (e.g., PE, MI, meningitis, ectopic pregnancy)
- Most likely: highest posterior probability given all clinical data
- Possible: plausible but lower probability or less urgent.
-
Apply LR to update probabilities — for key clinical findings and test results: post-test odds = pre-test odds × LR+/−. Use published LR values (EvidenceAlerts, DynaMed); a LR+10 with 10% pre-test probability gives 53% post-test probability.
-
Identify the discriminating workup — order tests that have the highest LR+ for the top "must-not-miss" and most-likely diagnoses; avoid ordering tests that won't change management regardless of result.
-
Apply diagnostic thresholds — test when: uncertainty is high enough to warrant testing (>test threshold) but not high enough to treat without confirmation (below treatment threshold); treat without testing when probability is above treatment threshold in time-sensitive conditions.
-
Iteratively update after each result — each finding (positive or negative) updates probabilities; re-rank the differential with each new piece of information. Do not anchor to the initial most-likely diagnosis if contradictory evidence accumulates.
Rules
- Always include the most dangerous diagnosis on the differential, even if probability is low — missing aortic dissection at 2% probability is more consequential than missing tension headache at 40%.
- Premature closure is the most dangerous diagnostic error — explicitly ask "What else could this be?" after reaching a leading diagnosis.
- Treat the patient, not the diagnosis — if clinical status deteriorates, widen the differential rather than explaining deterioration as a complication of the current diagnosis.
- Negative test results rule out diagnoses only if test sensitivity is high enough — a negative D-dimer rules out PE only when pre-test probability is low; a negative D-dimer with high pre-test probability does not.
Common Mistakes
- Anchoring — fixing on the first plausible diagnosis and seeking confirmatory evidence while ignoring contradictory findings; corrected by explicitly generating alternatives before ordering tests.
- Availability bias — over-weighting diagnoses seen recently or memorably; corrected by using base rate and systematic differential generation rather than pattern recall alone.
- Overordering without a diagnostic strategy — ordering a "shotgun" panel without knowing which results would change management generates noise, cost, and false positives without improving diagnostic accuracy.
- Forgetting zebras in high-risk contexts — rare diagnoses are rare in primary care but common in referral centers and in patients who have already been evaluated; context shifts the prior.
When NOT to Use
- In a time-critical emergency where pattern recognition and immediate treatment take precedence over comprehensive differential generation (treat shock first, refine diagnosis after stabilization)
- For straightforward, high-specificity presentations where the clinical picture is pathognomonic (e.g., classic shingles rash — treat, do not generate a broad differential)
- As a standalone AI exercise for individual patient medical decisions — differential diagnosis for real patients requires licensed clinical assessment, physical examination, and full access to the complete history
