Machine learning for modelling the health impacts of extreme heat: A comprehensive literature review

• Machine learning (ML) is increasingly used in environmental and health sciences. • We performed a literature review of current ML applications in heat-health studies. • Studies were from high-income countries and mainly predicted one health outcome. • Random Forest was the most widely used, but did not systematically perform best. • Deep learning and global datasets should be leveraged for populations most at risk. Extreme heat ranks among the deadliest weather events globally. The number of heat-related deaths is expected to rise sharply as population ages and climate changes. In recent years, machine learning (ML) approaches have been increasingly used across a range of environmental and health fields, including heat-health studies. In this paper, we conducted a comprehensive literature review of the current ML applications for modelling the human health impacts of extreme heat. We searched for relevant scientific articles published in English in PubMed, Scopus and Web of Science databases from their inception to the search date of December 20, 2024. After screening titles, abstracts and full-texts, 25 papers were included in this review. We found that most of the studies were conducted in high-income countries such as Japan, Canada, South Korea or the United States. The studies primarily modelled a single health outcome, including all-cause mortality or heat-related illness, in the general population. The main predictors of health outcome were maximum and mean air temperature, followed by relative humidity, and various temporal and socio-demographic variables. The most commonly used approach was Random Forest. Results were mixed regarding the optimal algorithm and most important predictors. This review highlighted the strengths and limitations of current ML applications in heat-health studies. We propose recommendations to help guide the future development of these approaches to reduce the heat-related health burden globally. Future research should 1) study multiple health endpoints at the individual level in vulnerable populations, 2) leverage deep learning with spatiotemporal representations of environmental predictors, and 3) use data from multiple locations at a high spatial resolution to provide insights in data-scarce regions.