A Multimodal Perceived Stress Classification Framework Using Wearable Physiological Sensors

Abstract

Mental stress is a common condition that poses serious health risks, but proper management can greatly improve quality of life. We propose a robust multimodal framework for perceived stress classification using data from forty subjects collected via three physiological modalities: electroencephalography (EEG), galvanic skin response (GSR), and photoplethysmography (PPG). Unlike most existing studies that focus on single modalities and binary classification, our framework addresses both two- and three-class perceived stress problems through multimodal fusion. Data was acquired over three minutes in an open-eye condition, and stress levels were assessed using the Perceived Stress Scale to assign labels. Time-domain features were extracted from GSR and PPG signals, while frequency-domain features were extracted from EEG. A frequency band selection algorithm identified the theta band as optimal for stress classification, and a wrapper-based feature selection method was applied to derive an effective multimodal feature set. Stress classification was performed with three classifiers utilizing features from all modalities. Among these classifiers, a significant accuracy (95% for two classes and 77.5% for three classes) was achieved using multilayer perceptron. The fusion of features from multiple modalities improves perceived stress classification, and our method, based on wearable sensors, is feasible for out-of-lab applications.