Feature Augmented Hybrid CNN for Stress Recognition Using Wrist-based Photoplethysmography Sensor.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Stress significantly impacts mental and physical health, with increased global prevalence due to events like the COVID-19 pandemic.
• Continuous stress monitoring is crucial for timely intervention and management.
• Wrist-worn smartwatches with photoplethysmography (PPG) sensors offer a convenient platform for physiological signal monitoring.

Purpose of the Study:

• To develop and evaluate a novel hybrid Convolutional Neural Network (H-CNN) classifier for stress detection.
• To leverage wrist-based PPG signals (Blood Volume Pulse - BVP) for stress detection applicable to consumer-grade smartwatches.
• To combine hand-crafted features with automatically extracted features for improved stress classification accuracy.

Main Methods:

• Utilized Blood Volume Pulse (BVP) signals from wrist-based PPG sensors.
• Developed a hybrid CNN (H-CNN) model integrating classical machine learning features and deep learning (CNN) extracted features.
• Evaluated the H-CNN model on the benchmark WESAD dataset for both 3-class (Baseline vs. Stress vs. Amusement) and 2-class (Stress vs. Non-stress) classification.

Main Results:

• The H-CNN model demonstrated superior performance over traditional classifiers and standard CNNs.
• For 3-class classification, H-CNN achieved ≈5% higher accuracy and ≈10% higher macro F1 score than traditional methods.
• For 2-class classification, H-CNN showed ≈3% higher accuracy and ≈3% higher macro F1 score compared to traditional methods.

Conclusions:

• The proposed H-CNN classifier effectively detects stress using PPG signals from wrist-worn devices.
• Hybrid feature extraction in H-CNN enhances stress detection accuracy, offering a promising approach for wearable health technology.
• This method provides a viable solution for continuous, non-invasive stress monitoring in consumer smartwatches.