Multi-region models built with machine and deep learning for predicting several heat-related health outcomes

• Daily summer mortality and morbidity predicted for general and elderly populations. • Machine/deep learning models trained with weather, regional and temporal predictors. • Deep learning performed better globally and across regions than other ML models. • Ensemble tree-based models were more sensitive to weather variables and heatwaves. • Models provided new insights on heat-related health impacts in the two populations. As a result of climate change, populations worldwide will be exposed to more heat episodes. To ensure a sustainable future, cutting-edge tools must be developed to predict the health effects of heat and limit its consequences. However, current research has mainly focused on one health outcome in a single city/region, thus providing limited knowledge to improve society's resilience to extreme heat. In this study, a machine learning (ML) framework is introduced to predict several heat-related health outcomes in multiple regions simultaneously, using the province of Quebec (Canada) as a case study. Five ML models including penalized regression, ensemble tree-based models and deep neural networks were considered and compared. Models were trained to predict these health outcomes using various meteorological, regional and temporal predictors across all regions. Our results showed that deep learning models were the most promising, with out-of-sample R 2 of >60 % for most of the studied health outcomes. However, ensemble tree-based approaches also had the best performance for some health outcomes, and were more sensitive to weather variables and to heatwaves. By introducing novel ML-based tools for predicting heat risks in several regions, this study can guide climate change adaptation and help cities and society to become more healthy, resilient and sustainable.