Ensemble Models Predict Multilevel Health Care Usage

A Singapore research team (2020–2022) developed and temporally validated multiclass stacking ensemble models to predict inpatient length of stay and emergency department visits for patients with type 2 diabetes. Trained on 108,886 and validated on 111,004 patients, boosted-tree ensembles achieved multiclass AUCs of 0.6877 (LOS) and 0.7601 (ED) and estimated a simulated SGD $152 million annual cost reduction for one model. The study evaluates predictive performance and real-world economic impact for population health planning.