Integrating healthcare system context to improve risk prediction and reduce disparities among dual-eligible Medicare-Medicaid beneficiaries

Sanjay Basu

Abstract: Medicare-Medicaid dual-eligible beneficiaries comprise less than 15% of enrollees but account for disproportionate spending. Current models used to predict future spending rely on patient demographics and diagnoses, potentially missing healthcare system context (provider availability, care fragmentation, facility characteristics). We estimated the degree to which healthcare context data improve risk prediction and what portion of risk and risk disparities among racial/ethnic groups is predicted by such data. Adding healthcare system context to patient data improved model performance from R-squared=0.454 to 0.615 for prospective spending prediction (a 35% improvement). The sensitivity of predicting acute care visits improved from 25.0% to 33.8%, while maintaining specificity (greater than 97%). Context-enriched XGBoost models outperformed conventional regression approaches; system-context features comprised approximately 25% of top predictors. Contextual factors were associated with attenuated rather than amplified risk for most beneficiaries, challenging assumptions that system-level context mainly increases risk for vulnerable populations. Healthcare delivery context substantially improves prediction of future spending among dual-eligible beneficiaries. These findings suggest risk adjustment incorporating context should account for differential racial impacts.