Stress Classification Using Photoplethysmogram-Based Spatial and Frequency Domain Images.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Stress classification is challenging due to individual variability.
• Generic models offer limited accuracy, while personalized models require extensive data and frequent updates.
• Photoplethysmogram (PPG) signals contain valuable physiological data for stress assessment.

Purpose of the Study:

• To develop a robust stress classification approach using PPG signals.
• To explore the efficacy of spatial and frequency domain image transformations of inter-beat intervals (IBI).
• To compare the performance of person-specific, generic, and calibrated generic classification models.

Main Methods:

• Inter-beat intervals (IBIs) were extracted from PPG signals.
• IBIs were converted into spatial and frequency domain images.
• Convolutional Neural Networks (CNNs) were employed for stress classification.
• Three model types were evaluated: person-specific, generic, and calibrated generic.

Main Results:

• Person-specific models achieved near-perfect accuracy (up to 100%) on training, validation, and test sets.
• Calibrated generic models, incorporating a small percentage of external data, significantly outperformed generic models.
• Frequency domain images demonstrated high effectiveness in stress classification across all model types.

Conclusions:

• Transforming PPG-derived IBI data into frequency domain images is a promising strategy for stress classification.
• Calibrated generic models offer a practical solution balancing accuracy and data requirements.
• This approach enhances the reliability and efficiency of individual stress state monitoring.