Personal comfort models in long-term monitoring using physiological data from wearable sensors

Abstract

Personal Comfort Models (PCMs) propose a new approach for human-centric comfort studies overcoming the one-size-fits-all of the conventional models. This research addresses the development of PCMs based on a seven-month long-term monitoring campaign including continuous environmental and physiological data collection through wearables and daily survey submission about subjects’ sensations. To tackle the influence of subjects’ environmental exposure history, time series of environmental data of different durations were used to predict individuals’ perception via Machine Learning models with Support Vector Machine and Random Forest methods. The accuracy and F1-score values of seven different PCMs were confronted for each subject and for the whole group (nine people). The number of datapoints per subject and their answers’ consistency during time affected the models’ accuracy, and the inclusion of physiological signals improved the models’ performance. When considering the whole dataset, the comfort model accuracy decreases supporting that individual subjectivity have an important impact in the environmental perception prediction.