Physiological Data Analysis for Stress Detection utilizing 1D-CNNs

Abstract

A majority of people experience stressful situations in their life. These situations differ, causing varying physiological and psychological effects on individual, many of them harmful. Because of these unhealthy responses, improving the detection of stress is crucial for effcient reacting and therefore healing these consequences of stress. Fortunately, technology has evolved to offer various different non-invasive wearable devices for measuring different biosignals, from which these stressful situations can be identifed and analyzed.

This study aims to explore the capabilities of 1D-CNNs on binary classifcation (BC) problem of classifying wearable sensor data, and the differences in performance metrics of these data modalities. The data points are classifed either stress or non-stress. For each data modality, an independent 1D-CNN is trained and tested. During testing phase, different performance metrics are calculated, and fnally they are compared between different data modalities. The data used for analysis and training the networks is the publicly available Wearable Stress and Affect Detection (WESAD) dataset. The BC-problem results from WESAD are used as indicative baseline.

Based on this study, EDA and TEMP are the best data modalities for 1D-CNNs, EDA being clearly the best. Although the accuracies did not surpass 90 %, by experimenting more with the hyperparameters the results may yet be improved. Utilizing 1D-CNNs in stress detection clearly has signifcance, and further experimentation is needed for better results. Collecting physiological data that satisfes quality requirements is a challenging task especially for stress detection, hence improving the classifcation methods in speed and accuracy is crucial for benefting from current and future stress data.