Bias mitigation in clinical AI: audit methods for health equity
Systematic methodology for identifying, quantifying, and mitigating bias in clinical AI systems, with focus on health equity outcomes across populations.
Authors
J. Moreau, GALEN, 29 contributors
Published
2030
Citations
185
Overview
Clinical AI systems are trained on historical patient data that reflects existing healthcare disparities. This results in AI systems that perpetuate or amplify health inequities. This research develops practical audit and mitigation frameworks tested across 12 healthcare systems.
Methodology
Bias audits of 34 clinical AI systems across 12 healthcare networks. Demographic stratification of performance metrics across race, gender, age, socioeconomic status, and geographic location. Root cause analysis of performance disparities. Mitigation testing (data augmentation, algorithmic fairness techniques, training set rebalancing). Health outcome tracking.
Key Findings
Clinical AI systems show significant performance disparities even when trained with demographic awareness: average performance gaps across major demographic groups are 12-18% (measured as diagnostic accuracy variation). These gaps correlate with underrepresentation of minority populations in training data: systems trained with <10% minority representation show 2.2x larger disparities.
The source of bias is not primarily model architecture but data representation: models trained on representative data (matching population demographics) show 73% reduction in performance disparities even without explicit fairness constraints. This suggests training data curation is more impactful than algorithmic fairness techniques.
Simple approaches to bias mitigation (demographic rebalancing, audit oversight) reduce health disparities by 48-62%. More sophisticated approaches (algorithmic fairness constraints, outcome-based fairness definitions) add only 8-12% additional improvement, suggesting simple interventions capture most of the benefit.
Sustained equity improvement requires organizational commitment: systems receiving annual audits and retraining based on equity metrics maintain improvements; those audited once show performance regression within 12 months. Healthcare organizations embedding equity audits into quality assurance processes achieve sustained equity outcomes.
Impact & Application
Audit frameworks adopted by 18 healthcare systems affecting care for 8M+ patients. Shapes FDA guidance on clinical AI validation. Reduces documented health disparities by 35-50% across adopting systems.
Contributors
Lead: Dr. Jean-Paul Moreau (Computational Medicine school). Collaborators from Johns Hopkins, Mayo Clinic, Massachusetts General Hospital. Advisors from NIH and CDC.