Learning to Combine: An Interpretable Adaptive Ensemble for Lassa Fever Incidence Prediction

Lassa fever is an acute viral hemorrhagic disease endemic to West African countries, including Nigeria, Sierra Leone, Liberia, and Guinea. Nigeria experiences recurrent outbreaks that impose public health, economic, and social burdens, particularly in endemic states such as Ondo and Edo. To the best of our knowledge, despite recent advances in AI, no study has investigated adaptive ensemble learning specifically for Lassa fever incidence prediction. Hence, this study proposes an Interpretable Adaptive Ensemble Learning (IALE) framework, which combines SARIMA with an adaptively selected recurrent neural network (LSTM or GRU) through a Gradient Boosting stacking ensemble while employing SHAP to provide interpretable explanations of the final incidence predictions. Weekly confirmed Lassa fever incidence data was obtained from the Nigeria Centre for Disease Control and Prevention (NCDC). 2018–2023 data was used for training and 2024 for testing. The adaptive ensemble model combined the optimal SARIMA model with LSTM/GRU whose hyperparameters were selected using random search with early stopping to avoid overfitting. The better performing deep learning model was selected for each state and combined with SARIMA using Gradient Boosting meta-learner. SHAP was applied to interpret the contribution of the base predictions to the final predictions. Model was evaluated using RMSE, MAE and R². The Adaptive Ensemble outperformed all individual models for both states with RMSE, MAE and R² of 0.668, 0.535 and 0.99 respectively in Ondo State and 0.417, 0.339 and 0.996 respectively in Edo State. These findings show that the adaptive ensemble effectively harnesses the strength of both statistical and deep learning models for Lassa fever incidence prediction.