Stress recognition identifying relevant facial action units through explainable artificial intelligence and machine...

Area of Science:

• Psychophysiology
• Machine Learning
• Computer Vision

Background:

• Facial cues and expressions offer insights into stress levels.
• Facial Action Units (AUs) model facial muscle movements for stress assessment.
• This study explores automatic acute stress recognition using AUs.

Purpose of the Study:

• Investigate automatic acute stress recognition based on facial Action Units (AUs).
• Utilize conventional Machine Learning and Deep Learning techniques for stress detection.
• Identify relevant AU combinations for stress condition classification.

Main Methods:

• Developed a new dataset with 58 participants and 11 stress/non-stress tasks.
• Employed computational feature selection for robust AU subset identification.
• Integrated AUs with Machine Learning and Deep Learning using Layer-Wise Relevance Propagation for analysis.

Main Results:

• Acute stress significantly increased AU presence and intensity compared to neutral states.
• Identified and ranked the most relevant AU combinations for specific stress types.
• Achieved a mean classification accuracy greater than 93% for stress vs. neutral conditions.

Conclusions:

• Specific AU combinations are key indicators of acute stress conditions.
• These AU combinations improve the separability between neutral and stress states.
• Facial Action Units offer a reliable method for automatic stress recognition.