Investigating Functional Data Analysis for Wearable Physiological Sensor Data in Stress Evaluation

Abstract

Quantifying an individual's stress level objectively is crucial for supporting personalized health monitoring and intervention. While traditional methods typically demand a clinical setting, capturing physiological signals with wearable devices presents a promising, real-time, non-invasive alternative that can be performed remotely. Much of the literature on this topic focuses on distinguishing between stress exposure and non-exposure, with limited applicability of findings to real-world settings. In this paper, we reformulate the problem as a regression task, focus solely on stress exposure observations, and evaluate the use of Functional Data Analysis (FDA) methods to enhance the information extracted from physiological signals. We apply scalar-on-function regression and functional clustering to WESAD, a public dataset containing signals from wearable devices and psychometric questionnaires that we used as a ground truth for stress. We compare FDA results with results obtained by some widely used techniques working on features extracted from signals rather than the signals themselves (linear regression, Random Forest, and clustering based on features distance). The comparison demonstrates the efficacy of FDA in extracting richer information by capturing the signals' variation over time. This advantage extends to the interpretation of the association between signals and stress, enabling new insights on how such association changes with different stressing activities. Moreover, we identify both instances where non-functional techniques are sufficient to capture groups and trends, as well as instances where FDA is key to capture overtime patterns linked to stress levels.