Individualized Stress Mobile Sensing Using Self-Supervised Pre-Training.

Area of Science:

• Mobile sensing and wearable technology
• Machine learning for health
• Digital health and personalized medicine

Background:

• Stress is a major health concern, and real-time prediction via wearables can enable timely interventions.
• Electrodermal activity (EDA) is a common biosignal for stress, but machine learning prediction faces challenges like subjective labels and sparse data.
• Model personalization, training separate models for each user, is a promising approach to address these challenges.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning (SSL) for pre-training neural networks for personalized stress prediction.
• To enable accurate stress detection with minimal user-labeled data through personalized model training.
• To improve the feasibility of mobile sensing for complex and subjective health outcomes like stress.

Main Methods:

• Utilized a one-dimensional convolutional neural network (1D CNN) pre-trained with self-supervised learning (SSL) on biosignal data.
• Employed model personalization, training individual stress prediction models for each user.
• Evaluated the method on the Wearable Stress and Affect Detection (WESAD) dataset, comparing SSL pre-trained models against supervised baselines.

Main Results:

• SSL pre-training enabled personalized models to outperform supervised baselines.
• Models trained with SSL required less than 30% of the labeled data to achieve equivalent performance compared to non-personalized SSL models.
• Demonstrated the effectiveness of learned embeddings from SSL for personalized stress prediction.

Conclusions:

• Personalized stress prediction using SSL pre-training is effective and data-efficient.
• This approach facilitates precision health systems requiring minimal user annotation.
• Enables advanced mobile sensing for subjective and heterogeneous health outcomes like stress.