District-level XGBoost forecasting model for dengue fever in Sri Lanka using climate and epidemiological data

Abstract

Dengue is still a major public health issue in Sri Lanka, with seasonal epidemics putting pressure on the health system. Reliable localised early warning systems are necessary for effective disease control and resource planning. This research suggests a machine learning model for forecasting dengue cases monthly at the district level in Sri Lanka. The study aims to develop an enhanced predictive framework that can accurately forecast dengue incidence, enabling more accurate and targeted public health enhancement. The main goal is to use an improved XGBoost model to predict dengue cases in each district and to create useful features such as past trends and time-based patterns that help capture how the disease spreads over time. 25 individual forecasting models were developed for each Regional Director of Health Services (RDHS) and were subdivided using the Extreme Gradient Boosting (XGBoost) algorithm. Each of the models has been trained on time-series data from January 2018 to May 2024. The main predictors were district-level historical monthly dengue case counts and monthly precipitation and average temperature from January 2018 to May 2024. Features were also created to include temporal lags of up to six months for climate and epidemiological data, along with cyclical time-based features to identify seasonality. The methodology integrates epidemiological data, detailed weather parameters (rainfall and temperature) from the NASA POWER API and newly engineered features, including extended lagged variables, advanced temporal indicators (trend components, moving averages). Outlier capping was also applied to robust the model against extreme values. Performance was tracked on the basis of Root Mean Squared Error (RMSE) and R-squared (R2). An Overall Test R2 of 0.686 and an overall Test RMSE of 0.6557 were observed across all the districts. Monaragala, Trincomalee, and Badulla exhibited high prediction accuracies with R2 values of 0.7801, 0.7414, and 0.7349, respectively. This research shows the potential to build strong and localised early warning systems for dengue in areas where the disease is common. The outcomes have public health policy implications, enabling targeted interventions to reduce the impact of dengue epidemics, and contribution to improved public health outcomes in Sri Lanka.