Contrastive Self-Supervised Learning for Stress Detection from ECG Data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Electrocardiogram (ECG)-based stress assessment is gaining traction due to its non-invasive nature and data accessibility.
• Current models predominantly use supervised learning, requiring extensive manual data annotation.
• Self-supervised learning (SSL), particularly contrastive SSL, remains underexplored for ECG stress assessment.

Purpose of the Study:

• To propose and evaluate a contrastive SSL model for ECG-based stress assessment.
• To investigate the efficacy of the SimCLR framework for stress detection using unlabeled ECG data.
• To compare the performance of the proposed model against state-of-the-art (SOTA) SSL methods.

Main Methods:

• Development of a contrastive self-supervised learning model for ECG stress assessment, based on the SimCLR framework.
• Training and testing the model on two established ECG-based stress assessment datasets (WESAD and RML).
• Comparative analysis against existing SOTA ECG-based SSL models.

Main Results:

• The proposed contrastive SSL model achieved a 9% accuracy improvement on the WESAD dataset.
• The model demonstrated a 3.7% accuracy improvement on the RML dataset compared to SOTA.
• The results indicate the potential of contrastive SSL for accurate, data-efficient stress assessment.

Conclusions:

• Contrastive SSL, using the SimCLR framework, is effective for ECG-based stress assessment.
• This approach enhances accuracy and reduces reliance on manually annotated data.
• The findings support the development of advanced wearable health monitoring and stress management tools.