Electrocardiogram-Based Mental Stress Detection Amid Everyday Activities Using Machine Learning: Model Development...

Area of Science:

• Biomedical Engineering
• Data Science
• Cardiology

Background:

• Frequent stress impacts health, necessitating continuous monitoring.
• Electrocardiograms (ECG) offer noninvasive, continuous stress biomarkers via wearables.
• Distinguishing mental stress (MS) from daily activities using ECG and machine learning (ML) is complex.

Purpose of the Study:

• Evaluate ML models for distinguishing MS from non-stress states.
• Assess model generalizability across new stressors and participants.
• Test model robustness for lightweight wearable suitability (lower sampling rates, fewer features).

Main Methods:

• Utilized a comprehensive ECG dataset (1000 Hz, 127 participants) with diverse stressors and activities.
• Extracted 55 features from 30-second ECG windows; trained logistic regression (LR) and extreme gradient boosting (XGBoost) models.
• Performed leave-one-stressor-out analysis, downsampling, feature reduction, and window sensitivity tests.

Main Results:

• XGBoost and LR models achieved comparable performance (AUROC ~0.74, AUPRC ~0.71).
• Models demonstrated robustness to downsampling and feature reduction (>93% performance with 10 features).
• Specificity for differentiating stress from moderate physical activity was poor (LR: 0.444, XGBoost: 0.418).

Conclusions:

• ML models effectively detect MS with high sensitivity and robustness to reduced data.
• Model generalization varied by stressor, with limited transfer to social-evaluative stress.
• Distinguishing ECG stress markers from physical exertion poses a significant limitation for single-sensor approaches.