Forecasting Disease Progression to Prevent High-Cost Acute Episodes in Vulnerable Medicare Advantage Populations

Abstract

Rising healthcare expenditures in Medicare Advantage (MA) populations are disproportionately driven by high-cost

acute episodes (HCAEs) among vulnerable subgroups with multiple chronic conditions. Traditional reactive care

models fail to anticipate disease progression, leading to preventable hospitalizations and emergency department

overuse. This study addresses the gap in predictive analytics tailored specifically to clinically vulnerable MA

enrollees by developing and evaluating a machine learning-based forecasting framework for HCAEs. Using a

retrospective cohort design, we analyzed five years of electronic health records and claims data (N = 45,000) from a

large MA plan serving a socioeconomically diverse region. The primary methodology integrated gradient boosting

machines (XGBoost) with time-varying risk markers to forecast 30-day HCAE risk. Key findings demonstrate that

our model achieved an area under the curve (AUC) of 0.89 (95% CI: 0.87–0.91), outperforming traditional risk

adjustment methods. Most influential predictors included labile biomarker trajectories, recent acute care utilization,

and social determinants of health. Implications suggest that embedding such forecasts into clinical workflows could

reduce avoidable HCAEs by up to 34%, lowering per-member-per-month costs while improving health equity. This

research contributes a validated predictive framework for proactive, precision-oriented population health

management.