Identifying those most at risk during heatwaves: AI limits and acting on what we already know

Abstract Background Heatwaves are an increasing public health threat, driven by the increasing frequency and severity of extreme heat events due to climate change. Protecting individuals most vulnerable to heat-related harm is a major public health priority. While there is growing interest in using electronic health records and artificial intelligence (AI) to predict individual risk, the utility of such approaches remains uncertain. Methods A time-stratified case-crossover study design was used to examine associations between temperature and mortality, using national primary care records in England. Conditional logistic regression was employed to estimate the effects of individual-level clinical and socio-environmental risk factors on heat-related mortality. Additionally, a machine learning approach, random forest, was used to assess the feasibility of predicting individual risk of death during heat events based on primary care data. Results Several clinical and socioenvironmental risk factors were consistently associated with increased mortality during heatwaves, including cardiovascular and respiratory diseases, mental health conditions, use of specific medications, and socio-demographic factors. However, some unexpected patterns were also observed, suggesting that context plays a critical role in shaping vulnerability. The random forest model performed poorly in predicting individual-level heat-related mortality, with predictive accuracy close to random chance highlighting the limitations of current data and methods for individual risk stratification and complexity of heat vulnerability. Conclusions The findings suggest that while predictive models have academic appeal, their practical utility in this context is limited. Rather than continuing to invest in increasingly complex data-driven models, public health efforts should focus on applying existing evidence to identify and support at-risk groups.