A Bi-LSTM Prediction Model Integrated with GIS for Spatiotemporal Malaria Endemicity Mapping and Early Warning in Indonesia

Abstract

Malaria continues to be a major public health challenge in Indonesia, particularly in eastern provinces where transmission patterns are influenced by climatic variability, geographical heterogeneity, and historical incidence trends. This study proposes an integrated spatio-temporal malaria forecasting and early warning framework by combining Bidirectional Long Short-Term Memory (Bi-LSTM), Geographic Information Systems (GIS), and SHapley Additive exPlanations (SHAP). Monthly malaria incidence, climate variables, population data, and provincial spatial data from 12 endemic provinces in Indonesia during 2014–2025 were used. The data were preprocessed through incidence-rate conversion, outlier handling, log transformation, Min-Max normalization, and six-month sliding window segmentation. The proposed Bi-LSTM model was assesed using RMSE, sMAPE, and R², and compared againts Naive Forecasting, SARIMA, and Simple LSTM baselines. The model attained optimol global performance, with an RMSE of 0.0522, sMAPE of 18.39%, and R² of 0.9553. The provincial analysis shows good performance throughout most regions, including high-burden areas like Papua and West Papua, however a decline in relative accuracy was observed in West Nusa Tenggara due to near-zero incidence rates. SHAP analysis revealed that historical malaria incidence was the primary predictor, whereas rainfall emerged as the most significant climatic variable. GIS-based forecasting showed spatial patterns aligned with malaria epidemiology in Indonesia, with Papua exhibiting the gratest predicted incidence in December 2025. These findings demonstrate that the Bi-LSTM–GIS–SHAP framework can support malaria endemicity mapping, interpretable forecasting, and province-level early warning for targeted public health interventions.