A machine-learning Approach for Stress Detection Using Wearable Sensors in Free-living Environments

Abstract

AbstractStress is a psychological condition due to the body’s response to a challenging situation. If a person is exposed to prolonged periods and various forms of stress, their physical and mental health can be negatively affected, leading to chronic health problems. It is important to detect stress in its initial stages to prevent psychological and physical stress-related issues. Thus, there must be alternative and effective solutions for spontaneous stress monitoring. Wearable sensors are one of the most prominent solutions, given their capacity to collect data continuously in real-time. Wearable sensors, among others, have been widely used to bridge existing gaps in stress monitoring thanks to their non-intrusive nature. Besides, they can continuously monitor vital signs, e.g., heart rate and activity. Yet, most existing works have focused on data acquired in controlled settings. To this end, our study aims to propose a machine learning-based approach for detecting the onsets of stress in a free-living environment using wearable sensors. The authors utilized the SWEET dataset collected from 240 subjects via electrocardiography (ECG), skin temperature (ST), and skin conductance (SC). In this work, four machine learning models were tested on this data set consisting of 240 subjects, namely K-Nearest Neighbors (KNN), Support vector classification (SVC), Decision Tree (DT), and Random Forest (RF). These models were trained and tested on four data scenarios. The K-Nearest Neighbor (KNN) model had the highest accuracy of 98%, while the other models also performed satisfactorily.