Global Prediction of Dengue Incidence Using an Explainable Artificial Intelligence-Driven ConvLSTM Integrating...

Dengue fever is a rapidly expanding vector-borne disease that poses significant global epidemiological and public health challenges. Accurate and interpretable forecasting is essential for timely interventions, yet most models overlook spatiotemporal, sex-specific, and country-level heterogeneity in disease dynamics. This study aimed to develop a robust explainable AI (XAI) framework to predict dengue incidence globally and identify key environmental, health system, and socio-economic drivers.

A Convolutional Long Short-Term Memory (ConvLSTM) network was applied to predict dengue incidence across 118 countries from 2000 to 2021. The data set included total, male, and female dengue incodence alongside 20 climatic, environmental, health system, and socio-economic predictors. The model was trained using data from 2000 to 2018 and tested on 2019-2021. Model performance was evaluated using RMSE, MAE, R², and adjusted R². Feature contributions were assessed through multiple XAI approaches: SHAP values, permutation importance, ±50% perturbation sensitivity perturbations, integrated gradients (IG), and layer-wise relevance propagation (LRP).

ConvLSTM achieved the best predictive performance (R² = 0.7731), demonstrating suitability for national-level public health planning. Sex-specific analysis revealed that annual freshwater withdrawals (SHAP: 44.37%; IG: 0.41; LRP: 0.38) dominated male dengue incidence, whereas hospital bed density had a greater influence for females (SHAP: 31.86%; IG: 0.34; LRP: 0.32). Temperature anomalies contributed consistently to both sexes (SHAP: 11.51%; IG: 0.18; LRP: 0.17). Country-level contributions highlighted electricity access (India: 97.35%; Bangladesh: 89.62%) and UHC coverage (Bangladesh: 8.33%) as key socio-economic determinants, with environmental and healthcare factors such as community health resources (Afghanistan: 35.42%; Brazil: 9.00%) further shaping sex-specific patterns. Sensitivity analysis indicated dengue incidence varied from -65% to +91% under ±50% predictor perturbations, underscoring model responsiveness and targeted intervention potential.

By integrating SHAP, IG, and LRP, the ConvLSTM-XAI framework provides transparent, sex-aware, and country-specific dengue forecasts. The results support targeted, climate-resilient, and equitable dengue control strategies.