New AI model predicts indoor heat exposure during pregnancy with precision
Researchers developed a machine learning system that accurately forecasts daily indoor temperature and humidity in homes, potentially revolutionizing how public health agencies assess heat-related risks to pregnant women. The model analyzed nearly 1,000 homes and could help identify vulnerable populations before dangerous heat events occur.
Originaltitel: Prediction of daily home indoor temperature and relative humidity using a deep ensemble machine learning approach
• We developed indoor temperature and humidity models using ensemble machine learning • The study used a large dataset from 978 participants across 1,029 homes • Models included 56 predictors covering meteorology, building, and occupant factors • Models captured daily fluctuations well and showed adequate long-term performance • Models are applicable to future heat-related epidemiological studies. Available modelling frameworks for estimating indoor temperature (T) and relative humidity (RH) for epidemiological studies remain scarce. We developed a modelling framework to assess the daily mean indoor T and RH. We monitored indoor T and RH at 1,029 homes of 978 participants from the Barcelona Life Study Cohort (BiSC), Spain (2018-2021), for one week each during the first and third trimesters of pregnancy. We applied a Deep Ensemble Machine Learning (DEML) approach to predict the daily mean indoor T and RH throughout pregnancy, which integrated predictions from three base models: Random Forest, eXtreme Gradient Boosting, and Gradient Boosting Machine. The models incorporated a comprehensive set of 56 predictor variables, including meteorological conditions, building and neighborhood characteristics, and occupants’ sociodemographic and behavioral characteristics. We applied a long-term validation to assess model performance across pregnancy and a short-term validation to evaluate daily fluctuation capture. The DEML model achieved excellent performance in the short-term validation (T: R² = 0.978, MAD = 0.312°C; RH: R² = 0.894, MAD = 1.666%), with a good performance for indoor T (R² = 0.891, MAD = 0.717°C) and a moderate performance for RH (R² = 0.499, MAD = 3.591%) in the long-term validation. Feature importance analysis indicated that the previous one-day mean outdoor T and the same-day outdoor RH were the most influential predictors for indoor T and RH, respectively. The model reliably predicted indoor T and RH, highlighting its utility for future epidemiological studies on health impacts of indoor exposure.