Building and occupant characteristics as predictors of temperature-related health hazards in American homes

• Extreme temp. exposure often happens at home, but role of buildings is unclear • We built machine learning models to predict temperature-related illness in US homes • Compared model performance with different variables from US household survey data • Including building-related variables improves model accuracy, recall, and precision • Results aid public health planning to mitigate temperature-related health hazards Many cities and regions are making significant investments towards planning for extreme temperature and in particular extreme heat. A heat vulnerability index (HVI) is a metric to track spatial variation in extreme temperature risk to target mitigation interventions. Most HVIs focus on demographic characteristics, which generally relate to vulnerability, and lack information about the building stock, which mediate the occupant's exposure to extreme temperatures. In this study, we use the Energy Information Administration's (EIA) Residential Energy Consumption Survey (RECS) to estimate prevalence of temperature-related illness in the United States and develop machine learning models using climate, demographic, and building characteristics to predict them. Temperature-related illness affects approximately 2 million households annually, around 1% of the total population. The models we developed predict temperature-related illness with up to 85% accuracy. The most important feature is energy insecurity, which describes the household's ability to maintain and operate heating, ventilation, and air conditioning (HVAC) systems. Our results offer guidance for municipalities to improve data collection, enabling them to better identify at-risk households and strategize resources for short-term and long-term interventions.