Symptom Attribution Bias (Wrong Diagnosis Due to Anchoring)

Issue: Model Anchors on Initial Symptom; Misses True Diagnosis Because Anchored to Wrong Hypothesis

Frequency: Very Common

Symptoms

• Model gets one symptom; overcommits to diagnosis
• Ignores contradictory symptoms
• Fails to revise diagnosis when new info arrives
• “Tunnel vision” diagnosis (ignores context)

Root Cause Models learn statistical associations (chest pain → heart disease, 90% of time). But 10% of chest pain is not cardiac. Model overweights early evidence; doesn’t properly update with Bayesian reasoning. Anchoring bias (psychological effect) replicated in AI.

Example

Scenario: Chest pain triage
Patient: Middle-aged male + chest pain
Model probability: "Cardiac event" = 85%
Actually: Pulmonary embolism (rare for this demographic but present)
Other symptoms: Leg swelling, shortness of breath (consistent with PE, not MI)
Model: Ignores leg symptoms (low weight in training); sticks with MI diagnosis
Impact: Patient gets wrong treatment; life-threatening delay

Key Statistics

• Sensitivity to contradictory evidence: 20-40% (models slow to change)
• Bayesian update deficit: Models underweight new evidence vs. prior

Mitigation Strategies

1. Explainable Reasoning: Show reasoning for diagnosis; allow clinician to challenge
2. Alternative Hypotheses: Generate top-3 diagnoses, not just 1
3. Evidence Weighting: Require certain contradictory symptoms to override diagnosis
4. Conditional Probabilities: Model P(symptoms|diagnosis) properly, not just P(diagnosis|symptoms)

Metrics

• Sensitivity to contradictory evidence (does model update?)
• Diagnostic accuracy when multiple hypotheses plausible
• Calibration of confidence

Alerts

• Confidence >90% with contradictory evidence → Warn clinician

References

• Diagnostic Reasoning in AI
• Bayesian Updating in Medical AI