Optimizing Temporal Segmentation of Multi-Modal Non-EEG Signals for Human Stress Analysis

Prolonged or severe exposure to stress is the underlying etiology of several diseases. Stress is a defensive mechanism that follows multiple stages of physiological responses, culminating in adaptation or coping of the body. However, frequent occurrence of stress can lead to fatal mental and physical consequences. Hence incessant stress monitoring in subjects, even during routine tasks, is vital. More critical is the identification of stress during its early onset, when the physiological response is at its peak. Quantifying stress using EEG based technology has been prevalent for the past several decades. But due to its lack of portability and the comparatively early stages of single channel acquisition devices, use of EEG based devices for continuous remote stress monitoring is quite impractical. The most suitable alternative is human stress analysis through non- EEG physiological signals acquired using wearable devices like smart watches and electronic finger rings. This paper investigates continuous stress monitoring using features from non-EEG signals such as acceleration (movement, gravity, noise), temperature, electrodermal activity, arterial oxygen level, and heart rate. Data from 20 healthy subj ects was collected during Relaxation, Physical stress, Cognitive stress, and Emotional stress states. Sample epochs were temporally segmented into windows representing the onset, sustenance, and offset of stress. Eighty-one linear and non-linear measures were extracted from each windowed epoch. Stress state identification was performed using Logistic Regression and Random Forest models, with performance evaluated via stratified 5-fold cross- validation. Findings indicate stress impacts are greatest at onset, with L-minute and 2-minute windows yielding higher classification accuracies than 5-minute windows. Optimizing temporal segmentation can reduce stress identification time by 60%.