Contrastive Pretraining for Stress Detection with Multimodal Wearable Sensor Data and Surveys

Area of Science:

• Computational neuroscience
• Machine learning for health

Background:

• Stress significantly impacts mental and physical health, necessitating early detection.
• Wearable sensors and other data sources are used for real-world stress monitoring.
• Self-supervised learning methods are increasingly used for stress detection due to high labeling costs.

Purpose of the Study:

• To investigate effective self-supervised learning models for stress detection using multimodal data.
• To address the understudied area of combining time series and tabular features in self-supervised stress detection.
• To develop a method for training models with varying data granularity and limited stress labels.

Main Methods:

• Developed a self-supervised multimodal learning framework for stress detection.
• Integrated time series physiological signals with tabular data (demographics, traits, context).
• Evaluated model performance with different data granularities and limited labels.

Main Results:

• The proposed self-supervised multimodal approach demonstrates effectiveness in stress detection.
• Combining time series and tabular features enhances stress monitoring capabilities.
• The method shows promise for practical applications with limited labeled data.

Conclusions:

• Self-supervised multimodal learning is a viable approach for stress detection.
• Integrating diverse data types improves the accuracy and robustness of stress monitoring.
• This method offers a promising solution for accessible and effective stress management tools.