Multimodal wearable sensor-based stress detection: machine learning pipeline with systematic feature selection and...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Sensor Technology

Background:

• Growing awareness of stress-related health issues necessitates advanced, non-invasive stress detection.
• Wearable sensors offer a promising avenue for continuous physiological monitoring.
• The role of feature selection in multimodal stress detection models requires further investigation.

Purpose of the Study:

• To develop and evaluate a machine learning pipeline for mental stress detection using multimodal sensor data.
• To assess the impact of systematic feature selection on model performance.
• To compare multimodal sensor fusion against unimodal approaches for stress classification.

Main Methods:

• Collected physiological data (electrodermal activity (EDA), electrocardiography (ECG), electroencephalography (EEG)) from 17 participants.
• Implemented a machine learning pipeline including data preprocessing, feature extraction, and classification.
• Applied four feature selection methods (ANOVA, Chi2, Kruskal-Wallis, MRMR) and validated externally on the SRAD dataset.

Main Results:

• Multimodal sensor fusion increased classification accuracy by 12.9%, achieving up to 95.9%.
• Feature selection provided an average accuracy gain of 4.8%, with Chi-squared (Chi2) yielding the highest performance.
• Identified key biomarkers: ECG (median, mean, root-mean-square), EEG (beta-to-alpha ratio, relative alpha power), and EDA (mean, sum phasic activity).

Conclusions:

• Systematic feature selection is crucial for optimizing multimodal sensor-based mental stress detection.
• The developed pipeline enhances accuracy, robustness, and interpretability of stress detection systems.
• Findings underscore the potential of integrated physiological sensing and machine learning for mental health monitoring.