An interpretable machine learning approach to multimodal stress detection in a simulated office environment

Area of Science:

• Human-computer interaction
• Occupational health
• Machine learning

Background:

• Work-related stress significantly impacts employee health and productivity.
• Current stress detection methods often lack real-world applicability.
• Automated, unobtrusive stress detection can enable timely interventions.

Purpose of the Study:

• To develop and evaluate a machine learning methodology for stress detection in a realistic office environment.
• To investigate the efficacy of multimodal data, including behavioral and physiological signals, for stress detection.
• To identify key behavioral indicators of stress through interpretable machine learning models.

Main Methods:

• Utilized multimodal data (mouse, keyboard, heart rate variability) from 90 participants in a simulated office.
• Employed machine learning models (SVM, Random Forests, Gradient Boosting) with 10-fold cross-validation.
• Applied SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• Gradient Boosting models using mouse and keyboard features achieved high F1 scores (0.625 for stress, 0.631 for arousal, 0.775 for valence).
• Behavioral features (mouse movements, typing patterns) proved more effective than heart rate variability for stress detection in this context.
• SHAP analysis identified specific mouse and typing behaviors linked to different stress levels.

Conclusions:

• This study bridges the gap between controlled lab experiments and real-world stress detection.
• Mouse and keyboard interaction data offer a viable, unobtrusive approach to detecting stress in office settings.
• Findings support the development of personalized, interpretable ML systems for real-time stress management.