Eliminating Individual Bias to Improve Stress Detection from Multimodal Physiological Data

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Psychophysiology

Background:

• Stress monitoring is crucial for mental wellbeing and early disorder detection.
• Non-invasive physiological signals offer a promising avenue for objective stress assessment.
• Existing methods often rely solely on Electroencephalogram (EEG) features, potentially limiting accuracy.

Purpose of the Study:

• To investigate the efficacy of fusing multiple physiological signals (EEG, PPG, GSR) for enhanced stress detection.
• To develop a subject-specific stress classification model by mitigating individual biases.
• To identify optimal physiological features for stress recognition using feature selection techniques.

Main Methods:

• Early fusion of features from EEG, Photoplethysmogram (PPG), and Galvanic Skin Response (GSR) signals.
• Maximum Relevance Minimum Redundancy (mRMR) for physiological feature selection.
• Subject clustering based on self-reported Valence, Arousal, and Dominance to address inter-subject variability.

Main Results:

• Early fusion of EEG, PPG, and GSR features improved stress detection F-score from 0.61 (EEG only) to 0.64.
• The proposed clustering technique before fusion and classification further boosted the mean F-score to 0.69.
• Leave-One-Subject-Out cross-validation demonstrated the effectiveness of the subject-specific approach.

Conclusions:

• Multi-modal physiological signal fusion significantly enhances stress detection accuracy compared to EEG alone.
• A clustering-based approach effectively reduces subject-specific bias, leading to more robust stress classification.
• This work provides a promising framework for developing more accurate and personalized stress monitoring systems.