A multi-modal deep learning approach for stress detection using physiological signals: integrating time and frequency...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Occupational Health

Background:

• Stress detection is crucial in high-pressure occupations like nursing.
• Wearable devices offer potential for continuous physiological data collection.
• Existing methods may lack robustness in real-world, intermittent data scenarios.

Purpose of the Study:

• To develop a multimodal deep learning-based stress detection method (MMFD-SD).
• To utilize intermittently collected physiological signals (accelerometer, EDA, HR, skin temperature).
• To validate the method in the nursing profession as a representative high-pressure occupation.

Main Methods:

• A multimodal deep learning framework integrating time-domain and frequency-domain features.
• Application of data augmentation (sliding window, jittering) and Synthetic Minority Over-sampling Technique (SMOTE).
• Customized Convolutional Neural Networks (CNNs) architecture for feature processing and classification.

Main Results:

• The MMFD-SD method achieved 91.00% accuracy and 0.91 F1-score.
• Significant improvement in accuracy and robustness compared to traditional machine learning models.
• Ablation studies confirmed the critical role of integrating both feature domains.

Conclusions:

• The MMFD-SD model offers an accurate and robust solution for stress detection using integrated physiological features.
• The method is suitable for occupational settings with intermittent data collection.
• Future work includes exploring real-time detection and enhanced model generalization.