Random forest based thermal comfort prediction from gender-specific physiological parameters using wearable sensing technology

Abstract

Prior knowledge of occupants’ thermal comfort can facilitate informed control decision of ambient thermal-conditioning in a building environment. This paper investigates the possibility to predict human thermal state (Comfort/Discomfort) from the information of physiological parameters. As gender difference has been widely linked with thermal comfort perception, a four-fold objective is adopted: first, investigating gender differences in subjective thermal perception; second, investigating gender differences in physiological response under different thermal states; third, identifying those physiological features that have the potential to predict thermal state, and fourth, establishing a data-driven thermal state prediction model for each gender group using the identified features. Human subject experiments were conducted, during which five physiological responses (hand skin temperature STh, hand skin conductance SCh, pulse rate PR, blood oxygen saturation SpO2, blood pressure BP) and four subjective responses (thermal comfort, thermal preference, humidity sensation, airflow sensation) were recorded in conjunction with a thermal sensation survey while environmental conditions varied from cold/cool to neutral levels (18°C−27∘C). Additionally, derivative features namely change rate (FOG) and mean squared gradient (MSG) of each physiological parameter were examined. Rigorous statistical analysis and subsequent predictive modeling utilizing Random Forest algorithm were implemented. Results demonstrate significant gender difference in several subjective and physiological responses. The features identified for males (SpO2-MSG/STh/STh-FOG/STh-MSG/SCh/SCh-FOG/SCh-MSG) and females (STh-FOG/STh-MSG/SCh/SCh-MSG/PR/PR-MSG) could accurately predict 92.86% and 94.29% of thermal states, respectively. This study indicates that the thermal state of a person can be identified by monitoring physiological parameters from non-intrusive body locations using wearable sensing technology.