Early-warning prediction of visceral leishmaniasis mortality using a multivariate STL–deep learning hybrid approach on 20 years of monthly time series

Introduction Visceral leishmaniasis (VL) is a preventable disease, but continues to cause mortality in Sudan, with transmission dynamics and potentially fatal outcomes strongly affected by local environmental conditions. Methods This research presents an innovative hybrid forecasting framework that amalgamates Seasonal-Trend decomposition using Loess (STL) with four sophisticated models: Gaussian Process Regression (GPR), Long Short-Term Memory (LSTM), Temporal Pattern Attention-LSTM (TPA-LSTM), and Light Gradient Boosting Machine (LightGBM), to forecast climate-induced multivariate VL mortality in Gedaref State, Sudan. Twenty years of monthly time series data from 2002 to 2022 were used, integrating VL mortality counts with meteorological variables such as precipitation, temperature, and relative humidity. The model’s performance was evaluated using MAE, RMSE, MAPE, R 2 , Willmott Index, and PBIAS. Results Among the models, STL-LightGBM exhibited the best predictive accuracy ( R 2 = 0.9491), whereas the deep learning approaches inadequately captured non-linearities, long-term dependencies, and seasonal changes. In this work, we concentrate on mortality prediction, hence directly contributing to a large research gap that has not been tackled by other works, which have been focused on the prediction of VL incidence. Discussion This proposed system has great potential in being an early-warning tool, which could be used to predict death surges and the seasonal variation, contribute by distributing pharmaceuticals and diagnostic devices, and help prepare rural health systems. These findings demonstrate the great potential of hybrid decomposition-learning models in the prediction of NTDs in regionally specific, resource-limited and climate-dependent regions.