Minimizing Racial Algorithmic Bias when Predicting Electronic Health Record Data Completeness

Area of Science:

• Health Informatics
• Biostatistics
• Health Services Research

Background:

• Electronic health record (EHR) continuity algorithms previously showed suboptimal performance in racially diverse populations.
• Improving the accuracy of EHR continuity prediction is essential for equitable healthcare research.

Purpose of the Study:

• To enhance an EHR continuity algorithm's performance by optimizing its race modeling strategy.
• To reduce algorithmic bias in EHR continuity predictions for racial minorities.

Main Methods:

• A claims-linked EHR dataset was randomly divided into training (70%) and testing (30%) sets.
• Models were developed with and without race interactions and race-specific models, using cross-validated LASSO for predictor selection.
• Performance was compared using the area under the receiver operating curve (AUC) on a held-out Medicaid-linked EHR validation dataset.

Main Results:

• In the validation set, incorporating race-interaction terms improved model performance in Black (AUC 0.821 vs. 0.812) and other non-White race (AUC 0.828 vs. 0.812) subgroups.
• Race-specific models showed performance comparable to models with race-interaction terms within racial subgroups.
• The race interactions model reduced misclassification of comparative effectiveness research (CER) variables by 2-3 fold for individuals with high predicted EHR continuity.

Conclusions:

• Inclusion of race-interaction terms significantly improved EHR continuity algorithm performance in racial subgroups.
• This optimized algorithm has the potential to mitigate algorithmic bias against racial minorities in EHR data analysis.
• Enhancing EHR continuity prediction accuracy is vital for equitable comparative effectiveness research.