Machine learning early warning for urban heat risk with CMIP6 projections.

Extreme heat (EH) has become an increasing threat to urban resilience, especially in rapidly urbanizing arid regions where environmental pressures interact with social vulnerability. Although heat related risks are rising, existing early warning approaches often fail to jointly account for socio demographic vulnerability, concurrent environmental drivers such as air pollution, and delayed exposure effects. This study addresses this gap by developing and validating a transferable urban heat early warning framework, using Arizona as a pilot case. We integrate alternative Heat Vulnerability Index (HVI) constructions for 2010-2020 with temperature, air quality indicators, and CO2 to project heat related mortality risks over 2025-2035 under CMIP6 SSP1.2.6 and SSP5.8.5 scenarios. A range of statistical and machine learning models is evaluated using a chronological holdout split and time-series cross-validation. Results show that the factor-analysis-based HVI produced the most stable downstream performance, and the final selected Ridge model achieved a holdout R2 of 0.5288 and a mean time-series cross-validated R2 of 0.5230. Projection results indicate a sustained upward shift in mortality risk under both SSP pathways, with estimates remaining sensitive to assumptions about future HVI and AQI trajectories. To enhance policy relevance, model outputs are translated into an operational Severe Heat Risk Level tool. By emphasizing computational efficiency, open source data, lag-aware co-exposure modeling, and explicit uncertainty assessment, the proposed framework offers a scalable and evidence based solution to support urban planners and decision makers in designing targeted heat adaptation strategies.