On-site measurement and machine learning prediction of age-specific outdoor thermal and humidity comfort in urban microclimates in summer

The accurate assessment of human thermal and humidity comfort in dynamic outdoor urban environments remains a critical challenge. On-site measurements with advanced machine learning (ML) are advancing. The lack of comprehensive field datasets hinders robust predictions across diverse age groups. This study conducted 12-day outdoor experiments with 12 participants (6 young and 6 elderly) to gather microclimate, physiological, and voting data from diverse urban environments and age groups. Three ML algorithms—backpropagation neural network (NN), random forest (RF), and support vector machine (SVM)—were evaluated. A Bayesian-optimized, SMOTE-Tomek method further improved performance. Spatiotemporal microclimatic variations in urban areas induced significant (p < 0.05) physiological responses in electrocardiogram (ECG) and electroencephalogram (EEG) activity. Preliminary findings suggest age-based divergences in outdoor thermal comfort. Young subjects tended to feel hotter and showed greater physiological stress, while the elderly appeared to perceive greater humidity yet remained more comfortable. Based on statistical results, eight input and three output parameters were selected, with emphasis on humidity sensation votes. RF performed best for the thermal comfort vote (TCV) and humidity sensation vote (HSV). After optimization, the RF model achieved final accuracies of 71 % for TCV (n = 394), 75 % for HSV (dry) (n = 196), and 83 % for HSV (wet) (n = 330). Including age as a predictor substantially improved accuracy, by up to 11 %. A post-hoc power analysis (power = 0.82) confirmed statistical adequacy for detecting large effects. This exploratory age-related model may inform more adaptive, age-specific urban planning pending validation with larger cohorts.