A Reproducible and Explainable Machine Learning Framework for Post-2023 Dengue Early Warning in Bangladesh Using Secondary Surveillance and Meteorological Data
Document Type : Original Article
Authors
- 1 Hazi Muhammad Mohsin College, Chittagong, Bangladesh.
- 2 Sabujbagh College, Dhaka, Bangladesh.
- 3 Rajendrapur Cantonment Public School & College, Gazipur, Bangladesh
- 4 Armed Police Battalion Public School & College, Nishindara, Bogura, Bangladesh
Abstract
Background: Dengue fever has emerged as a serious public health hazard in Bangladesh, with the 2023 epidemic exceeding the previous 23-year total of 321,179 confirmed cases and 1,705 deaths. However, despite the expanding use of machine learning in dengue prediction, substantial problems remain in post-2023 model validation, replication using publically available data, and model output interpretation.
Methods: An explainable multi linear (ML) framework was created using Bangladesh Meteorological Department climate data and Directorate General of Health Services (DGHS)/Institute of Epidemiology, Disease Control and Research (IEDCR) surveillance data (2010-2023), with a strict temporal holdout (2022-2023), and lagged meteorological predictors to compare eXtreme Gradient Boosting (XGBoost), Random Forest, Gradient Boosting, and Linear Regression, followed by SHapley Additive exPlanations (SHAP)-based explainability analysis.
Results: XGBoost outscored other models, including linear regression (R² = 0.412), with the best predictive performance (R² = 0.912, Root Mean Squared Error (RMSE) = 1,247 cases/month, Mean Absolute Error (MAE) = 876, Mean Absolute Percentage Error (MAPE) = 18.3%, and Area Under Curve (AUC) = 0.89). SHAP analysis identified two-month delay rainfall and one-month lagged relative humidity (with a ~80% RH threshold) as key predictors of dengue incidence. The mean temperature demonstrated a nonlinear relationship, with maximal risk in the 25-28°C region, which is Aedes aegypti's thermal niche. The model performed well overall; it overestimated the August-October 2023 peak due to its atypical character, but the 0-100 risk scoring system detected High and Very High risk periods with 95.8% accuracy and 100% sensitivity.
Conclusion: This study presents Bangladesh's first rigorously reproducible, SHAP-integrated early warning technique, which serves as a portable template for dengue surveillance in South and Southeast Asia based on the 2023 epidemic perspective.
Methods: An explainable multi linear (ML) framework was created using Bangladesh Meteorological Department climate data and Directorate General of Health Services (DGHS)/Institute of Epidemiology, Disease Control and Research (IEDCR) surveillance data (2010-2023), with a strict temporal holdout (2022-2023), and lagged meteorological predictors to compare eXtreme Gradient Boosting (XGBoost), Random Forest, Gradient Boosting, and Linear Regression, followed by SHapley Additive exPlanations (SHAP)-based explainability analysis.
Results: XGBoost outscored other models, including linear regression (R² = 0.412), with the best predictive performance (R² = 0.912, Root Mean Squared Error (RMSE) = 1,247 cases/month, Mean Absolute Error (MAE) = 876, Mean Absolute Percentage Error (MAPE) = 18.3%, and Area Under Curve (AUC) = 0.89). SHAP analysis identified two-month delay rainfall and one-month lagged relative humidity (with a ~80% RH threshold) as key predictors of dengue incidence. The mean temperature demonstrated a nonlinear relationship, with maximal risk in the 25-28°C region, which is Aedes aegypti's thermal niche. The model performed well overall; it overestimated the August-October 2023 peak due to its atypical character, but the 0-100 risk scoring system detected High and Very High risk periods with 95.8% accuracy and 100% sensitivity.
Conclusion: This study presents Bangladesh's first rigorously reproducible, SHAP-integrated early warning technique, which serves as a portable template for dengue surveillance in South and Southeast Asia based on the 2023 epidemic perspective.