Enhancing Hospital Inventory Forecasting Accuracy through Hybrid and Ensemble Learning Models

Abstract

Demand forecasting for medical and consumable supplies in healthcare institutions is a challenging problem due to irregular usage patterns, seasonality, sudden demand spikes, and data sparsity, and inaccurate forecasts may lead to stock-outs, excessive inventory costs, and disruptions in patient care. This study proposes an anomaly-aware, hybrid and ensemble-based forecasting and decision support framework for short-term hospital inventory demand prediction using real-world operational data obtained from a hospital inventory management system. The proposed approach integrates density-based anomaly detection, material-level behavioral feature extraction, supervised time series transformation, and a multi-model ensemble architecture combining linear models, tree-based methods, and boosting-based learners, with model selection and weighting performed via time-series cross-validation. To ensure operational robustness, a multi-layer fallback strategy incorporating classical exponential smoothing and conservative heuristics is employed for data-scarce scenarios, and an interpretable rule-based forecast confidence score together with an integrated ABC–XYZ segmentation scheme is used to directly link forecasts with inventory control policies. Experimental results on real hospital inventory data demonstrate that the proposed framework significantly improves forecasting stability and accuracy compared to single-model approaches, particularly for heterogeneous and irregular consumption patterns, while providing a practical, explainable, and operationally actionable solution for hospital inventory management.