Approximating Bayesian spatial relative risk of dengue in Bandung: A gradient boosting surrogate model with demographic and climatic features

Abstract

Dengue Hemorrhagic Fever (DHF), a disease transmitted by Aedes aegypti mosquitoes, remains a critical public health issue. Bandung is recorded as having one of the highest numbers of DHF cases in Indonesia, exhibiting a concerning increasing trend annually. This persistent rise indicates that the current control efforts for DHF have not been optimal. In this study, the relative risk of DHF transmission across subdistricts in Bandung is estimated using the Besag, York, and Mollié (BYM) model within a Bayesian inference framework. The BYM model is extended by incorporating additional covariates: demographic factors, specifically the proportion of residents with blood type O, and climatic factors, including temperature, rainfall, wind speed, and humidity. These additions enhance model performance by specifying prior distributions for parameters corresponding to each covariate. Furthermore, to assess the potential of machine learning as a scalable forecasting tool, XGBoost and LightGBM are applied as non-parametric surrogates to the complex BYM model, utilizing covariates and binary adjacency matrices to capture spatial dependencies. Results suggest that climatic and demographic variables drive the spatial pattern of DHF, with the proportion of the population having blood type O emerging as the dominant variable improving both Bayesian and machine learning model performance. After hyperparameter tuning, both models approximated DHF relative risk at different levels of fidelity, yielding RMSE values of 0.30 for XGBoost and 0.16 for LightGBM. These findings demonstrate that LightGBM in particular can effectively emulate Bayesian spatial patterns, serving as a deployable tool for rapid risk forecasting while reducing reliance on computationally intensive MCMC simulations required for standard inference.