Improving Fairness and Mitigating Bias in Multicenter Electronic Health Records Models to Predict Glaucoma Outcomes

Area of Science:

• Ophthalmic AI
• Clinical Informatics
• Health Equity

Background:

• Glaucoma progression prediction models are crucial for patient care.
• Bias in electronic health records (EHRs) can lead to unfair AI outcomes.
• Evaluating and mitigating bias in clinical AI is essential for generalizability.

Purpose of the Study:

• To assess bias mitigation techniques in glaucoma prediction models using multicenter EHR data.
• To introduce FairOdds-AUC, a novel metric for balancing AI performance and fairness.
• To ensure equitable AI deployment in ophthalmology.

Main Methods:

• A cohort of 50,656 glaucoma patients from seven US institutions was analyzed.
• Five model architectures were trained to predict surgical progression, with and without bias mitigation.
• FairOdds-AUC was developed to evaluate fairness (equalized odds) alongside performance (AUROC).

Main Results:

• Inprocessing methods (IPW, adversarial fairness) showed superior fairness-performance tradeoffs.
• IPW improved FairOdds-AUC for transformer and fully connected networks, maintaining discrimination.
• Postprocessing and preprocessing methods had variable results and weaker generalizability.

Conclusions:

• Inprocessing bias mitigation consistently enhanced fairness across diverse sites.
• FairOdds-AUC provides a flexible metric for evaluating fairness in clinical AI.
• Fairness evaluations and training are recommended for future ophthalmic AI.