Machine Learning Approaches for Malaria Forecasting Using Environmental Drivers: A Case Study in Tak Province, Thailand

Abstract

Malaria remains a significant public health challenge in Thailand's border provinces, were traditional reactive surveillance limits outbreak prevention capabilities. This study systematically evaluated six machine learning algorithms (Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Ridge regression, Elastic Net, Lasso, and XGBoost) for operational malaria forecasting at 1-4-week horizons in Tak Province, Thailand. Using 13 years of surveillance data (2012-2024, n=689 epidemiological weeks) integrated with satellite-derived environmental predictors (rainfall, temperature, soil moisture) processed via Google Earth Engine (GEE), models were trained using chronological partitioning and evaluated on 2024 holdout data using coefficient of determination (R²) and root mean square error (RMSE). Algorithm-specific optimal performance was identified: SVM achieved superior 1-2-week forecasting (R² = 0.744 and 0.687, RMSE = 14.6 and 16.2 cases/week), while KNN excelled at 3-4-week horizons (R² = 0.748 and 0.731, RMSE = 14.7 and 15.2 cases/week). Statistical significance testing with bootstrap confidence intervals confirmed genuine algorithmic advantages rather than random variation. Historical case features dominated predictive performance, while environmental variables provided complementary information. Models successfully tracked temporal patterns including the 2022-2023 transmission rebound. The satellite-based framework provides scalable solutions for resource-limited settings, with 1-4-week lead times enabling proactive intervention planning to support Thailand's malaria elimination objectives. This operational forecasting approach offers a replicable template for similar endemic contexts across Southeast Asia.