Evaluating the Reproducibility of Physiological Stress Detection Models.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Wearable sensors offer promising approaches for physiological stress detection.
• Significant challenges persist, notably the lack of reproducibility across studies.
• Existing models often fail to generalize across different devices, populations, or study protocols.

Purpose of the Study:

• To evaluate the reproducibility and validity of machine learning models for stress detection.
• To develop and assess novel methods for enhancing stress detection model performance.
• To establish a foundation for reliable stress detection in real-world conditions.

Main Methods:

• Analyzed data from 90 participants across four independent controlled studies.
• Utilized two different types of physiological sensors and varying study protocols.
• Evaluated model performance across studies and developed new methods, including a clustering approach for classification.

Main Results:

• Models showed improved performance when tested on data from different studies.
• Developed methods consistently increased stress-detection performance irrespective of device, sensor, or stressor type.
• A novel clustering approach outperformed traditional thresholding for cross-study classification.

Conclusions:

• The study provides critical insights into the reproducibility of stress detection methods.
• The developed methods offer a pathway to more robust and generalizable stress detection models.
• This work is a prerequisite for advancing stress detection in free-living conditions.