Forecasting dengue in Bangladesh using meteorological variables with a novel feature selection approach.

Dengue, a mosquito-borne viral disease, continues to pose severe risks to public health and economic stability in tropical and subtropical regions, particularly in developing nations like Bangladesh. The necessity for advanced forecasting mechanisms has never been more critical to enhance the effectiveness of vector control strategies and resource allocations. This study formulates a dynamic data pipeline to forecast dengue incidence based on 13 meteorological variables using a suite of state-of-the-art machine learning models and custom features engineering, achieving an accuracy of 84.02%, marking a substantial improvement over existing studies. A novel wrapper feature selection algorithm employing a custom objective function is proposed in this study, which significantly improves model accuracy by 12.63% and reduces the mean absolute percentage error by 70.82%. The custom objective function's output can be transformed to quantify the contribution of each variable to the target variable's variability, providing deeper insights into the workings of black box models. The study concludes that relative humidity is redundant in predicting dengue infection, while meteorological factors exhibit more significant short-term impacts compared to long-term and immediate impacts. Sunshine (hours) emerges as the meteorological factor with the most immediate impact, whereas precipitation is the most impactful predictor over both short-term (8-month lag) and long-term (26-30-month lag) periods. Forecasts for 2024 using the best-performing model predict a rise in dengue cases starting in May, peaking at 24,000 cases per month by August and persisting at high levels through October before declining to half by year-end. These findings offer critical insights into temporal climate effects on dengue transmission, aiding the development of effective forecasting systems.