Transferring healthcare risk prediction models between Medicaid populations: a transfer learning evaluation

Cross-state deployment of Medicaid risk prediction models is challenged by demographic, policy, and care-delivery differences that create domain shift. We evaluated transfer learning methods for predicting acute care utilisation between Washington (source; n = 20,744) and Virginia (target; n = 28,901) Medicaid populations enrolled in high-risk care management, where outcome prevalence differed markedly (9.4% vs 25.6%). Nine approaches were compared: source-only and target-only logistic regression, prototypical networks, domain-adversarial neural networks, causal transfer learning, TabTransformer, Enhanced MAML, a meta-ensemble, and a simple average ensemble. On the Virginia hold-out set, the meta-ensemble achieved the highest discrimination (AUC 0.728, 95% CI 0.691-0.764) and best calibration (Brier 0.193, 95% CI 0.180-0.207). Source-only transfer performed similarly (AUC 0.725), with no significant difference (p = 0.454), and both outperformed target-only logistic regression (AUC 0.628; p < 0.001). Enhanced MAML (AUC 0.677) did not improve over naive transfer. Post-hoc isotonic regression substantially improved calibration across models, underscoring the importance of prevalence adjustment. Fairness analyses showed lower race/ethnicity equalized odds differences for target-adapted models (0.132-0.157) than source-only transfer (0.897), though disparities persisted. Findings suggest simple transfer plus recalibration can match complex methods; broader validation is needed.