Advancing Heat-Related Impact Forecast Using Multivariate Deep Learning Models

Heatwaves pose an increasing threat to public health under climate change. Despite evidence that health systems in high-latitude countries are insufficiently prepared for extreme heat, few studies have investigated the state-of-the-art deep learning (DL) models to forecast heat-related morbidity at seasonal lead times. This study develops and evaluates a multivariate, multi-step impact-based forecasting framework across Sweden for predicting heat-related morbidity using Neural Basis Expansion Analysis for Time Series (N-BEATS) models. N-BEATS models are developed and tested under recursive and multi-input–multi-output (MIMO) multi-step forecast strategies and compared with statistical baselines (ARIMA, naïve seasonal) and classical DL model (Long Short-Term Memory (LSTM)). Forecasts are generated using morbidity counts alone and in combination with exogenous covariates (Heat Wave Index and the number of individuals with respiratory diseases) while local and global modeling approaches are examined.

Results show that N-BEATS with both covariate and local modelling strategy significantly outperforms all baseline models with the lowest MAE, RMSE, and MASE values. N-BEATS shows greater data efficiency with iteratively refined residuals through fully connected backcast and forecast stacked blocks compared to LSTM, particularly when there is an extreme morbidity peak. Individually trained local N-BEATS models are more effective than the cross-learning global N-BEATS, even with similar seasonal peaks and lower data quantity. Regional differences in climate, hydrology, and demographics could hinder the effectiveness of global models and underscore the importance of localized adaptation plans and measurements. Models may also underperform during unprecedented periods, such as during the COVID-19 pandemic in 2021. The underperformance may have resulted from disruptions in healthcare during COVID, behavioral change from seeking healthcare, and selected covariates didn’t capture healthcare system capacity. Future study could be improved by testing model performance to incorporate a covariate that reflects healthcare system capacity, such as service load to enhance model’s robustness to similar system level shock.

The study offers a concrete step toward operational impact-based early warning systems by enabling national agencies to anticipate heatwave burdens when a seasonal heatwave alert is issued. By coupling hazard forecasting with health impact prediction, this work supports the development of impact-based early warning systems tailored to the growing risks of extreme heatwaves. Integrating morbidity forecasts into heat-health action plans can support public health agencies in proactive resource allocation, risk communication, and preparedness planning.