Machine-learning framework for conditional estimation and scenario-based projection of the heat index for public health interventions

Background: Heatwaves are becoming more frequent and intense in Bangladesh, particularly in urban areas such as Dhaka, where the combined effects of extreme heat and humidity pose serious public health risks. Temperature-based warning systems frequently underestimate physiological heat stress. Methods: This study proposes a machine-learning framework to estimate and project the Heat Index (HI) using 3-hourly meteorological records from 2014 to 2023. Because HI is a deterministic function of air temperature and relative humidity, the framework performs conditional estimation of HI based on meteorological predictors rather than independently forecasting HI itself. Five models: ARIMAX, SARIMAX, Random Forest Regressor (RFR), XGBoost, and Long Short-Term Memory (LSTM), were trained using air temperature, relative humidity, and seasonal indicators as predictors. For scenario-based projections beyond 2023, future temperature and humidity were approximated using historical monthly averages, generating scenario-based HI projections that preserve seasonal and diurnal patterns. These projections represent climatological scenarios rather than true meteorological forecasts. Results: The Random Forest Regressor (RFR) achieved the highest conditional estimation accuracy, with the lowest RMSE (0.85°C) and highest R² (0.987) on the test set. Empirical 95% prediction intervals achieved 98.85% coverage, indicating slightly conservative uncertainty bounds. Scenario-based projections yielded mean HI values of 29.02°C (optimistic), 29.90°C (moderate), and 31.33°C (pessimistic). A substantial proportion of projected 3-hourly periods fall within the “Extreme Caution” category (32–41°C), indicating persistently elevated heat-stress exposure under climatological assumptions. Conclusion: The proposed framework demonstrates strong potential for generating high-resolution scenario-based HI projections by capturing nonlinear temporal dynamics and sub-daily variability. These findings can support scenario-based early warning systems and inform adaptive urban heat-management strategies in climate-vulnerable cities such as Dhaka, although results should be interpreted as conditional projections rather than deterministic forecasts. Unlike conventional HI studies, this framework translates meteorological inputs into high-resolution, operational heat-risk insights by modeling temporal persistence at sub-daily scales.