Bridging ensemble model and public health practice: an approach for refining understanding of seasonal dengue transmission patterns in Bangladesh.

backgroundDengue fever remains a persistent public health threat in Dhaka, Bangladesh, necessitating effective early warning systems to enable timely interventions and mitigate impacts. This study develops and evaluates modeling approaches to produce a projection of the 2026 dengue season in Dhaka, which may help inform public health preparedness.

methodsDaily dengue case data from Dhaka (January 1, 2020 - December 30, 2024; 1,826 observations) were obtained from a publicly available Kaggle repository. A comprehensive modeling framework was developed comparing Seasonal Auto-Regressive Integrated Moving Average (SARIMA) with machine learning models (Random Forest, XGBoost, Support Vector Regression). A weighted ensemble model was constructed using inverse RMSE weights derived from validation performance to combine the strengths of individual approaches. Models were trained on a chronological 80% split (2020-2023) and evaluated on a 20% hold-out validation set (2024) using RMSE, MAE, MASE, and R

resultsMachine learning models were shown to have higher predictive performance as compared to the time series model - SARIMA. Random Forest was the most accurate with the lowest (RMSE: 34.64, MAE: 19.14, MASE: 0.47) as well as the highest R-square (R

conclusionThe proposed weighted ensemble model provides a projection for the 2026 dengue season in Dhaka, offering potential early warning of the timing and scale of the high transmission period. This work adds importantly to the evidence base for dengue transmission in the region and helps to refine public health practitioners' understanding of the seasonal transmission pattern of dengue, helping inform their annual planning. This may support proactive public health planning, such as vector management in January-February and healthcare resource mobilization from March onward.