Predicting malaria incidence using meteorological, environmental, and spatial factors in Guinea-Bissau: a machine learning approach

Background Malaria remains a major public health problem in Guinea-Bissau, where transmission is strongly influenced by climatic and environmental conditions. Despite progress, its incidence exhibits significant seasonal and geographic variations. This study aimed to analyze the influence of meteorological factors on malaria incidence, evaluate the ability of machine learning models for trend prediction, and conduct a comparative evaluation against conventional time-series approaches across the health regions of Guinea-Bissau. Methods The study used 924 monthly observations collected across Guinea-Bissau’s 11 health regions between January 2016 and December 2022, considering 28 candidate variables spanning geographical, demographic, epidemiological, environmental, and climatic dimensions. Three machine learning models Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) were developed and benchmarked against conventional time-series models, namely SARIMA, ETS, and GLMM with Splines. Performance was evaluated using RMSE, MAE, and R², validated across temporal horizons up to 2022, and supported by residual diagnostics and spatial autocorrelation analyses (Moran's I). SHAP values were used to interpret variable importance. Results The XGBoost model achieved the best predictive performance, with an R² of 0.78 (95% CI: 0.52–0.97), alongside an RMSE of 2.27 (95% CI: 1.28–3.15) and an MAE of 1.74 (95% CI: 0.89–2.63) on the test set (2021), as well as strong validation metrics an R² of 0.91(95% CI: 0.82–0.98), an RMSE of 1.57 (95% CI: 0.87–2.24) and an MAE of 1.22 (95% CI: 0.65–1.85) on the 2022 horizon. Machine learning algorithms consistently outperformed conventional temporal models (SARIMA, ETS, and GLMM with Splines) in capturing non-linear relationships. The models maintained robust generalization without significant residual spatial autocorrelation. Conclusions Machine learning models, particularly XGBoost, significantly outperform conventional time-series approaches in predicting malaria dynamics in Guinea-Bissau. Integrating epidemiological, climatic, and environmental data improves the identification of high-risk periods, contributing to early warning systems and optimized control interventions.