Machine-learning detection of stress severity expressed on a continuous scale using acoustic, verbal, visual, and...

Area of Science:

• Psychophysiology
• Machine Learning
• Computational Social Science

Background:

• Early detection of acute stress is crucial for mitigating negative health outcomes.
• Machine learning algorithms show potential for stress monitoring.
• Continuous stress severity detection is challenging due to data quality demands.

Purpose of the Study:

• To detect laboratory-induced stress using multimodal data.
• To explore the performance of machine learning algorithms in stress severity detection.
• To identify challenges in multimodal stress detection research.

Main Methods:

• A machine learning algorithm was trained on multimodal data (visual, acoustic, verbal, physiological).
• Participants (n=42) underwent a Trier Social Stress Test, with stress levels self-reported at five time-points.
• Data collected included voice, facial expressions, and physiological measures.

Main Results:

• Participants reported minimal to moderate stress.
• The algorithm showed a weak association with self-reported stress (r² = .154).
• Classification into stressed/non-stressed categories yielded acceptable to good performance during the presentation task (accuracy up to .71).

Conclusions:

• Multimodal stress detection requires large, diverse sample sizes, which are difficult to obtain in lab settings.
• Automation may reduce resource demands but introduces technological challenges.
• Future research should focus on improving ground truth data and studying general populations and clinical groups.