Stress detection using wearable device data: a knowledge discovery and recurrent deep learning approach

Abstract

Increased attention to mental well-being in recent days has created a need for noninvasive and continuous methods of stress monitoring. Conventional assessment tools like EEG based systems are accurate although are clinical and unsuitable for day to day usage. A hybrid methodology combining knowledge discovery techniques with deep learning to analyze multimodal physiological signals collected from public datasets such asWearable Stress and Affect Detection (WESAD) provide accelerometer (ACC), skin temperature (ST), and heart rate (HR) data that are collected using consumer-grade wearables like smart wristbands. For the preprocessing part, signal cleaning, Z-score normalization, followed by time-frequency domain feature extraction has been brought out to identify stress relevant trends. The study implements a modified version of recurrent deep learning model currently called Modified DA-SKIP RNN to handle temporal dependencies and nonlinear meaningful patterns in the data stream,using Advanced multi-modal architectures combining per-channel embedding, hierarchical-attention, bidirectional GRU, and for personalized affective computing, subject adaptation has been performed using transfer learning procedure. This architecture is further refined with a Projection Layer for dimensionality compression and a Domain-Aware Normalization step that helps representing subject specific and contextual results leading to better generalization across individuals. The Modified DA-SKIP RNN achieves 99.18% accuracy for binary classification and 96.39% four-class accuracy under non-subject-independent evaluation, while subject-independent Leave-One-Subject-Out (LOSO) evaluation demonstrates 94.34% accuracy after a short calibration phase, confirming effective generalization to unseen individuals. The implications of this research is towards the developers, educators, and mental health practitioners for a easily accessible and data driven stress management tool. With the emerging focus on personal well being in recent times, this study will contribute to the real world implementation.