Optimizing Care Management Targeting: A Machine Learning Comparison of Regression, Classification, and Ranking Approaches for Identifying Future High-Cost Patients.

Effective care management programs depend on accurately identifying patients who are likely to account for a disproportionate share of future health care expenditures. Unlike traditional risk-adjustment models, which emphasize population-level predictive accuracy, care management targeting is fundamentally a top-k decision problem focused on maximizing expenditure capture. This study used 2000-2023 MEPS data to compare regression-, classification-, and learning-to-rank machine learning approaches for predicting next-year health care expenditures. Primary outcomes included precision-at-k, cost capture-at-k, weighted cost capture-at-k, and AUC. Models were benchmarked against a persistence-based floor baseline using prior-year expenditures and an oracle upper bound representing perfect future information. Classification-based models consistently achieved the strongest operational targeting performance. Neural network and XGBoost classifiers outperformed regression and ranking approaches in identifying future high-cost individuals and maximizing expenditure capture. Differences in cost capture were statistically significant across yearly evaluations. More complex modeling extensions provided little additional benefit. Prior-year utilization and expenditure variables contributed substantially to predictive accuracy; however, models retained approximately 80%-90% of full-model performance after removal of these features, indicating meaningful predictive signal from sociodemographic characteristics, chronic disease burden, functional limitations, and self-reported health status. For care management programs operating under fixed enrollment constraints, classification-based prediction approaches appear better aligned with operational objectives than regression or ranking. Additional performance gains are more likely to arise from richer clinical and longitudinal data than from increased model complexity alone. Future AI-enabled care management systems may further improve resource allocation by integrating cost-risk prediction with estimates of patient engagement and intervention impactability.