Multimodal Wearable Sensor-Based Stress Detection: Machine Learning Pipeline with Systematic Feature Selection and Key Biomarker Insights

The increasing awareness of stress-related health impacts has driven demand for accurate, non-invasive stress detection methods, particularly those leveraging wearable sensors. While multimodal sensing approaches have shown promise in enhancing mental stress assessment, the critical role of feature selection in optimizing model performance remains underexplored. This study presents a comprehensive machine learning pipeline for mental stress detection that integrates data preprocessing, feature extraction, systematic feature selection, and classification. Using data collected from 17 participants, we classified stress and relaxation states based on three physiological signals: electrodermal activity (EDA), electrocardiography (ECG), and electroencephalography (EEG). Multimodal sensor fusion was compared against unimodal approaches to assess performance improvements. To identify the most informative features and improve model accuracy, we applied four feature selection methods: Analysis of Variance (ANOVA), Chi-squared (Chi2), Kruskal-Wallis (KW), and Minimum Redundancy Maximum Relevance (MRMR). External validation was conducted using the public Stress Recognition in Automobile Drivers (SRAD) dataset. Our results demonstrated a 12.9% increase in classification accuracy using multimodal data, reaching up to 95.9%, with feature selection contributing an average gain of 4.8%. Among the methods, Chi2 consistently achieved the highest mean accuracy across various feature sets. Key biomarkers included ECG-based median, mean, and root-mean-square; EEG-based beta-to-alpha ratio and relative alpha power; and EDA-based mean and sum phasic activity. These findings highlight the importance of integrating systematic feature selection with multimodal sensor data to enhance the accuracy, robustness, and interpretability of mental stress detection systems.