Minimizing Racial Algorithmic Bias when Predicting Electronic Health Record Data Completeness

The previously developed algorithm for identifying subjects with high electronic health record (EHR)-continuity performed suboptimally in racially diverse populations. We aimed to improve the performance by optimizing the race modeling strategy. We randomly divided TriNetX claims-linked EHR dataset from 11 US-based healthcare organizations into training (70%) and testing data (30%) to develop and test models with and without race interactions and race-specific models. We held out a Medicaid-linked EHR dataset as validation data. Study subjects were ≥18 years with ≥365 days of continuous insurance enrollment overlapping an EHR encounter. We used cross-validated least absolute shrinkage and selection operator (LASSO) to select predictors of high EHR-continuity. We compared the model performance using area under receiver operating curve (AUC). There were 550,859, 236,089, and 65,956 subjects in the training, testing, and validation datasets, respectively. In the validation set, the introduction of race-interaction terms resulted in improved model performance in Black (AUC 0.821 vs. 0.812, P < 0.001) and other non-White race (AUC 0.828 vs. 0.812, P < 0.001) subgroups. The performance of the race-specific models did not differ substantially from that of the models with race-interaction terms in the race subgroups. Using the race interactions model, subjects in the top 50% of predicted EHR-continuity had 2-3-fold lesser misclassification of 40 comparative effectiveness research (CER) relevant variables. The inclusion of race-interaction terms improved model performance in the race subgroups. Using the EHR-continuity prediction algorithm with race-interaction terms can potentially reduce algorithmic bias for racial minorities.