Prospective Medicare Expenditure Forecasting via Deep Neural Regression with Robust Loss Minimization and Game-Theoretic Feature Attributions

Prospective forecasting of patient-level healthcare expenditures is foundational to the financial viability of Accountable Care Organizations (ACOs), Medicare Advantage plans, and value-based risk-sharing agreements. However, conventional actuarial workflows rely on retrospective claims reviews or linear risk-adjustment indices (e.g., standard CMS-HCC) that fail to capture compounding, non-linear interactions across multi-morbid chronic conditions. Furthermore, healthcare expenditure data exhibits pronounced right-skewness, fat tails, and zero-inflation, causing classical ordinary least squares and symmetric loss functions to exhibit severe instability. In this study, we present an end-to-end prospective expenditure forecasting and clinical explainability framework evaluated on the Centers for Medicare & Medicaid Services (CMS) 2008–2010 Data Entrepreneurs’ Synthetic Public Use Files (DE-SynPUF). To eliminate auto-regressive target leakage, we construct a 20-dimensional feature space restricted strictly to baseline patient demographics, 11 discrete chronic disease indicators, an aggregate comorbidity score, and non-monetary prior utilization proxies (annual inpatient claim volume and average length of stay). We formulate the prediction problem as a robust deep regression task using a three-layer Feedforward Artificial Neural Network (ANN) integrated with L2 weight regularization, batch normalization, dropout, and a logarithmic target reparameterization trained under Huber loss minimization. Across standardized 5-fold cross-validation, the neural network achieves a Mean Absolute Error (MAE) of $3,811.18, a Root Mean Squared Error (RMSE) of $5,940.91, an R² score of 0.5418, and a Root Mean Squared Logarithmic Error (RMSLE) of 0.6587, significantly outperforming traditional linear actuarial baselines. To bridge the interpretability gap that impedes clinical adoption, we couple the inference pipeline with cooperative game-theoretic feature attribution via Shapley Additive exPlanations (SHAP), translating high-dimensional neural representations into local, dollar-denominated waterfall decompositions ($). Finally, we demonstrate an interactive clinical decision-support deployment enabling sub-second inference and proactive 'what-if' scenario modeling for population health management. Keywords: Medicare Expenditures, CMS DE-SynPUF, Prospective Healthcare Risk Adjustment, CMS-HCC, Deep Neural Regression, Heavy-Tailed Distributions, Huber Loss Minimization, Logarithmic Reparameterization, Comorbidity Synergies, 5-Fold Cross-Validation, Shapley Additive exPlanations (SHAP), Accountable Care Organizations (ACOs), Clinical Decision Suppo