Accurately forecasting individual pharmacy costs is crucial for resource planning and policy decision-making in healthcare systems. This study proposes a set of scalable hybrid deep learning and ensemble models aimed at predicting the total pharmacy expenses for patients identified as high utilizers within the Medicaid rebate program. Leveraging the comprehensive dataset available from the Healthcare Cost Prediction Dataset, this work evaluates and compares multiple advanced regression models including Random Forest Regressor, XGBoost Regressor, CatBoost Regressor, and LightGBM Regressor.

To enhance predictive accuracy and capture non-linear temporal dependencies in patient health profiles, we further integrate a hybrid Autoencoder-LSTM model. The autoencoder serves to extract compact, noise-reduced representations of patient features, while the LSTM model captures temporal dynamics and sequential patterns. Our experimental results demonstrate that hybrid models consistently outperform traditional regression approaches in terms of RMSE and MAE metrics, especially for patients with highly variable medication histories. The proposed framework is designed to be scalable and adaptable to other insurance claims data, thus offering potential utility for state-level Medicaid programs and pharmaceutical budget planning. Additionally, the interpretability of tree-based models aids in identifying key drivers of pharmacy cost, contributing to transparent and explainable healthcare analytics.

Keywords: Pharmacy Cost Prediction, Medicaid Rebate Program, High Utilizers, Autoencoder-LSTM, Random Forest, XGBoost, CatBoost, LightGBM, Healthcare Analytics, Scalable Deep Learning.