Dynamic knowledge distillation for climate-driven diarrhoea incidence prediction

Diarrhoeal disease remains a significant public health burden in Vietnam, with incidence strongly influenced by climatic variability. Accurate short- and long-term forecasts are essential for effective early warning systems (EWS) to enable timely interventions. We propose a Dynamic Knowledge Distillation Ensemble (DKDE) framework for provincial-scale diarrhoea incidence forecasting that integrates local climate predictors with spatially informed neighbour-incidence data. DKDE dynamically selects the most relevant teacher models for each province and forecast horizon, adaptively combining their predictions into a compact student model through a data-driven distillation process. Using monthly diarrhoea surveillance data from 55 provinces (1997-2017) and corresponding climate records, we evaluate DKDE against state-of-the-art deep learning models (PatchTST, LSTM, CNN, Transformer variants) and the best-known method on the same dataset, the Dynamic Weighted Ensemble (DWE). Across six forecast horizons, DKDE consistently outperforms all baselines, achieving its largest gains in short-term prediction—reducing average 1-month RMSE from 16.74 (DWE) to 16.57—while maintaining superior accuracy at longer horizons. Comprehensive ranking and performance-gap analyses demonstrate DKDE's robustness and stability across spatial and temporal scales.