Exploring Unsupervised Machine Learning Classification Methods for Physiological Stress Detection.

Area of Science:

• Wearable health technology
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Wearable devices have advanced for health monitoring, including stress detection.
• Current wearable stress monitors predominantly use supervised learning, requiring subjective and often inaccurate self-reported stress labels.
• Labeling physiological signals for stress is challenging due to the limitations of point-in-time, subjective self-reporting questionnaires.

Purpose of the Study:

• To explore the feasibility of unsupervised learning clustering algorithms for wearable stress monitoring.
• To evaluate unsupervised classifiers as an alternative to supervised methods that require labeled data.
• To assess the potential for non-invasive, continuous, and robust stress detection using unsupervised methods.

Main Methods:

• Investigated unsupervised learning clustering algorithms: Affinity Propagation, BIRCH, K-mean, Mini-Batch K-mean, Mean Shift, DBSCAN, and OPTICS.
• Compared the performance of unsupervised classifiers against traditional supervised machine learning models.
• Utilized two publicly available datasets for evaluating classification results.

Main Results:

• Unsupervised learning classifiers demonstrated comparable classification performance to supervised methods.
• The study confirmed that unsupervised classifiers do not require perceived stress labels for stress level classification.
• Results indicate the viability of unsupervised approaches for stress monitoring.

Conclusions:

• Unsupervised learning holds significant potential for developing non-invasive, continuous, and robust wearable stress monitoring systems.
• This approach overcomes the limitations associated with subjective stress labeling in supervised methods.
• Unsupervised clustering offers a feasible and effective strategy for physiological and pathological stress detection.