Climate‐Driven Advanced Machine Learning Approach for Dengue Incidence Forecasting in Bangladesh

ABSTRACT Background and Aims Dengue fever has become a significant and an increasing public health menace in Bangladesh. Despite the abundance of research on the dengue outbreaks, the majority of the studies are limited by the short‐term time frame of research, a narrow scope, or to one modeling methodology. Thus, a gap in the existing knowledge on long‐term transmission processes, climatic predictors, and most importantly, the comparative validity of rival forecasting frameworks of dengue in Bangladesh still exists. The proposed study will fill these gaps by examining the trends of long‐term dengue and providing systematic comparison of the statistical and machine‐learning forecasting systems. Methods The study examined 14 years of monthly dengue incidence data (2010–2024) along with meteorological variables (temperature, precipitation and relative humidity). The patterns of transmission were considered and four models were used to assess predictive performance: Negative Binomial Regression (NBR), XGBoost, Long Short‐Term Memory (LSTM) networks and Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX). Out of sample validation was used to evaluate the models and test mean absolute error (MAE) was used to measure the forecasting performance of the models. Results The findings indicate the presence of a dual‐scale temporal dynamic consisting of a highly predictable seasonal cycle with its peak in the third quarter of each year with irregular inter‐annual variability on top of it. There was a close linkage between dengue transmission and a synergistic climatic envelope with ideal temperature (26°C–30°C), moderate levels of rainfall (200–600 mm), and high levels of humidity (> 75%). Machine‐learning models showed significant test overfitting even though they are complex. SARIMAX model was more robust and generalized to give the lowest test MAE (17.0) and narrow confidence interval. The long‐term forecasts point to the development of a stable hyperendemic situation, where almost all months of the 2025–2034 period will be at high‐risk and the cumulative burden will be more than 2.3 million cases. Conclusion This paper shows that the robustness of models is superior to the complexity of the algorithms in forecasting dengue in Bangladesh. The results underscore the urgent need to switch the reactive outbreak response mode to long‐term and proactive surveillance and enhanced preparedness of the health system.