Bayesian Geospatial Calibration of Reinforcement Learning for Malaria Transmission Control: Parameter Estimation

Malaria has been one of the major public health issues that has not been extensively addressed. Controlling the spread of infectious diseases in space and time requires robust adaptive policies that significantly account for heterogeneity, uncertainty, and optimal sequential decision-making. This study presents an innovative framework that integrates Bayesian spatiotemporal modeling with reinforcement learning (RL) with the 5D3 algorithm. The disease risk at location i and time t is modeled using a logistic regression with spatial random effects and Bayesian inference performed using the non-reversible Metropolis-Hastings algorithm, and the parameter estimates are used to calibrate a stochastic reinforcement learning environment via episodic parameter sampling. The study identified significant drivers of malaria risk: rainfall, temperature, secondary and tertiary levels of education, higher wealth index, female gender, treated nets, and spray repellents, while quantifying uncertainty via credible intervals. The spatial random effect captured unmeasured local heterogeneity, and the temporal effect accounted for seasonality, which is essential for reliable parameter estimation. Therefore, a reinforcement learning agent can learn optimal, spatially adaptive intervention policies under uncertainty, making the model suitable for public health decision-making where spatial heterogeneity and uncertainty are prominent. The proposed calibrated model within a policy-learning environment using posterior samples can be replicated to simulate realistic transmission scenarios for malaria and evaluate dynamic control strategies.