Summary
An ETH Zurich preprint used physiological and environmental signals from 20 participants to distinguish ambient temperature, humidity and physical exertion with 94.7% Random Forest accuracy. The exploratory study is an early step toward person-centred heat monitoring and requires testing in real-world settings.
A medRxiv preprint from ETH Zurich describes an exploratory approach to monitoring how individuals respond to heat. In a study involving 20 participants, a Random Forest model distinguished between ambient temperature, relative humidity and physical exertion with 94.7% accuracy using a combination of physiological and environmental signals.
The work is aimed at a gap in population-based heat indices. Such indices describe environmental conditions but do not capture why two people exposed to similar conditions may experience different levels of physiological strain. The researchers investigated whether digitally collected signals could help identify the source of heat stress for each person.
How the study combined physiological and environmental signals
The researchers conducted a single-centre, prospective, cross-sectional, cross-over, observational and exploratory study. The 20 participants were exposed to various personalised heat stressors. Heart rate, core temperature and close body temperature were assessed with wearable sensors, while local sweat rate was measured using technical absorbents.
These measurements were treated as potential digital biomarkers: digitally captured signals that can be analysed to characterise a physiological state or response. The study combined measurements from the body with information about the surrounding environment, rather than relying on either category alone.
The model was trained to distinguish three types of exposure identified in the study: changes in ambient temperature, changes in relative humidity and physical exertion. This is a classification task, meaning that the system attempted to identify the type of stressor associated with a set of measurements. The study did not evaluate a clinical alert system or a treatment for heat-related illness.
Model accuracy varied with the type of heat stressor
The Random Forest model distinguished the three heat stressors with 94.7% accuracy in the study. The variables that contributed most to the classification differed according to the stressor.
For ambient-temperature conditions, environmental measurements were the main predictors. Physical-exertion conditions were identified primarily through physiological measurements. Relative-humidity conditions showed contributions from both physiological and environmental signals.
That pattern supports the study’s central premise: heat monitoring may be more informative when it combines what is happening around a person with how the person’s body is responding. The authors describe the approach as a possible foundation for automated, laboratory-independent and person-centred monitoring during heat exposure.
The findings come from a small, single-centre preprint study and are best understood as a digital-biomarker exploration rather than a validated medical monitoring product. The authors state that further testing is needed in more complex and real-life settings. Such testing would be important for determining how consistently the signal combinations work beyond the personalised exposures used in this study.