Demographic Bias in Diagnosis & Treatment Recommendations

Issue: Model Biased Toward/Against Certain Demographics (Race, Gender, Age); Different Accuracy for Different Groups

Frequency: Very Common

Symptoms

• Accuracy 95% for demographic A, 65% for demographic B
• Same condition misdiagnosed more often in one group
• Recommendations differ by demographic (treatment disparity)
• Model trained on skewed population data

Root Cause Training data reflects historical biases in healthcare. Certain demographics underrepresented in training data (e.g., women underrepresented in cardiac disease research). Genetic/physiological differences between populations sometimes real, but often model captures training bias not biology.

Example

Scenario: Cardiovascular disease risk model
Trained on: 80% male, 20% female data
Model learns: "Chest pain in women = less urgent" (correlation from data)
Reality: Women's cardiac symptoms different; model dangerously underestimates risk
Result: Women receive lower-risk classification; delayed treatment
Impact: Disparity in outcomes; liability

Key Statistics

• Accuracy gap by demographic: 10-30% (worse for minorities)
• Underrepresentation: Female in cardiac studies 40% vs. 50% population

Mitigation Strategies

1. Representative Training Data: Ensure demographics represented proportionally
2. Stratified Evaluation: Test accuracy separately by demographic
3. Fairness Constraints: Enforce equal sensitivity/specificity across groups
4. Explainability: Show which features drive demographic disparities

Metrics

• Sensitivity/Specificity by demographic group
• Demographic parity (equal treatment across groups)
• Disparate impact ratio (should be >80% per legal standard)

Alerts

• Accuracy gap >10% between groups → P1 (fairness issue)

References

• Bias in Medical AI
• Fair Machine Learning in Healthcare