Related Experiment Video
Updated: Jun 20, 2025
12:51
Evaluation of Commercial-Off-The-Shelf Wrist Wearables to Estimate Stress on Students
Published on: June 16, 2018
7.5K
Wearable Data From Subjects Playing Super Mario, Taking University Exams, or Performing Physical Exercise Help Detect
Filippo Corponi1, Bryan M Li1,2, Gerard Anmella3,4,5,6
1School of Informatics, University of Edinburgh, Edinburgh, United Kingdom.
JMIR Mhealth and Uhealth
|July 17, 2024
Summary
Self-supervised learning (SSL) significantly improved mood disorder detection from wearable data by overcoming annotation limitations. The choice of pretraining task and data size are key to successful SSL in personal sensing.
Area of Science:
- Wearable technology and digital phenotyping
- Machine learning for healthcare
- Mental health monitoring
Background:
- Personal sensing with wearables offers a promising approach for monitoring mood disorders (MDs).
- Collecting and annotating wearable data is resource-intensive, limiting the application of supervised machine learning.
- This data bottleneck hinders the development of effective MD detection systems.
Purpose of the Study:
- To advance the detection of acute mood disorder episodes using wearable data by leveraging self-supervised learning (SSL).
- To overcome the data annotation bottleneck in personal sensing for mental health.
- To introduce a novel transformer architecture (E4mer) and a large-scale open-access dataset (E4SelfLearning) for SSL in this domain.
Main Methods:
- Utilized open-access Empatica E4 wristband datasets from various personal sensing tasks.
- Developed a preprocessing pipeline for on-/off-body detection, sleep/wake detection, and segmentation.
- Introduced E4SelfLearning, the largest open-access collection of E4 data, and a novel E4mer transformer architecture for SSL and supervised learning.
- Assessed SSL pretraining's effectiveness compared to fully supervised baselines for MD detection.
Main Results:
- SSL significantly outperformed fully supervised methods, achieving 81.23% correct classification of recording segments compared to 75.35% (E4mer) and 72.02% (XGBoost).
- SSL performance was strongly correlated with the pretraining surrogate task and the amount of unlabeled data available.
- The E4mer architecture demonstrated effectiveness for both SSL and supervised learning.
Conclusions:
- Self-supervised learning effectively overcomes the annotation bottleneck in personal sensing for mood disorder detection.
- The selection of the pretraining surrogate task and the volume of unlabeled data are critical factors for SSL success.
- The E4mer architecture and E4SelfLearning dataset are valuable resources to advance SSL research in personal sensing.
