Implementation of a Stress Biomarker and Development of a Deep Neural Network-Based Multi-Mental State Classification...

Area of Science:

• Biomedical Engineering
• Data Science
• Wearable Technology

Background:

• Stress monitoring is crucial for public health.
• Existing methods often rely on indirect measures or limited datasets.
• Wearable biosensors offer potential for continuous, real-world stress assessment.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting stress levels from large-scale biosignal data.
• To assess the model's ability to interpret physiological patterns associated with stress.
• To determine if a Transformer model can accurately reproduce stress indices derived from wearable devices.

Main Methods:

• Utilized ~137,000 longitudinal biosignal measurements from Sejong City residents.
• Integrated static machine learning (Random Forest, LightGBM) and deep learning (LSTM, Transformer) models.
• Employed SHapley Additive exPlanations (SHAP) and Transformer attention visualization for interpretability.

Main Results:

• The Transformer model achieved ~98% classification accuracy on the large dataset.
• The model successfully captured both short-term biosignal fluctuations and long-term temporal structures.
• Model interpretability methods quantified biomarker contributions and revealed temporal interactions.

Conclusions:

• Deep learning models, particularly Transformers, can effectively predict stress levels from wearable biosignal data.
• The study provides a methodological foundation for large-scale, wearable-based stress monitoring.
• Engineering feasibility for a comprehensive stress monitoring system using device-derived data is demonstrated.