Interpretable general thermal comfort model based on physiological data from wearable bio sensors: Light Gradient Boosting Machine (LightGBM) and SHapley Additive exPlanations (SHAP)

Abstract

This study aims to develop a general thermal comfort model using physiological signals obtained from wristband-type wearable biosensors. Accordingly, we constructed and evaluated supervised machine learning models by leveraging a diverse array of features extracted from physiological signals, including electrodermal activity (EDA), photoplethysmogram (PPG), and skin temperature (SKT). The model's performance was evaluated using data collected from 18 subjects across controlled experimental settings. Further, this study employed leave one subject out cross validation (LOSOCV) instead of the traditional k-fold CV to assess the model's generalizability to new subjects. Furthermore, SHapley Addictive exPlanation (SHAP) was incorporated to augment the interpretability and transparency of the model. The LightGBM model demonstrated a commendable test accuracy of 79.7% in distinguishing thermal preferences, namely, “want warmer,” “comfort,” and “want cooler.” These findings underscore the feasibility of employing wearable biosensors to evaluate occupants’ thermal comfort in real-world environments. This study makes a significant contribution to the literature by laying the groundwork for a broadly applicable method of continuous, objective, and noninvasive thermal comfort monitoring among building occupants. Considering previous challenges associated with personalized thermal comfort models due to individual variability, our study represents a pivotal step toward the development of a generalized thermal comfort model.