Biophysical versus machine learning models for predicting rectal and skin temperatures in older adults

Abstract

This study compares the efficacy of machine learning models to traditional biophysical models in predicting rectal (T_re) and skin (T_sk) temperatures of older adults (≥60 years) during prolonged heat exposure. Five machine learning models were trained on data using 4-fold cross validation from 162 day-long (8–9h) sessions involving 76 older adults across six environments, from thermoneutral to heatwave conditions. These models were compared to three biophysical models: the JOS-3 model, the Gagge two-node model, and an optimised two-node model. Our findings show that machine learning models, particularly ridge regression, outperformed biophysical models in prediction accuracy. The ridge regression model achieved a Root-Mean Squared Error (RMSE) of 0.27 °C for T_re, and 0.73 °C for T_sk. Among the best biophysical models, the optimised two-node model achieved an RMSE of 0.40 °C for T_re, while JOS-3 achieved an RMSE of 0.74 °C for T_sk. Of all models, ridge regression had the highest proportion of participants with T_re RMSEs within clinically meaningful thresholds at 70% (<0.3 °C) and the highest proportion for T_sk at 88% (<1.0 °C), tied with the JOS-3 model. Our results suggest machine learning models better capture the complex thermoregulatory responses of older adults during prolonged heat exposure. The study highlights machine learning models' potential for personalised heat risk assessments and real-time predictions. Future research should expand upon training datasets, incorporate more dynamic conditions, and validate models in real-world settings. Integrating these models into home-based monitoring systems or wearable devices could enhance heat management strategies for older adults.