Variational Instance-Adaptive Personalized Stress Recognition Based on Wearable Sensor Signals.

Area of Science:

• Physiological computing
• Machine learning for healthcare
• Affective computing

Background:

• Stress significantly impacts physical and mental health, necessitating accurate recognition for management.
• Inter-individual variability in stress responses challenges generic recognition models, causing performance degradation.

Purpose of the Study:

• To develop a personalized stress recognition framework, VIAStress, that overcomes limitations of generic models.
• To address distribution shifts caused by inter-individual variability in stress detection.

Main Methods:

• Proposed Variational Instance-Adaptive Stress recognition (VIAStress) framework using Domain Generalization (DG).
• Employed Variational Instance Adaptation (VIA) to model classifier parameters as instance-conditioned distributions for personalized adaptation.
• Integrated Multimodal Domain Generalization (MMDG) to enhance feature representation by fusing Photoplethysmography (PPG) and Electrodermal Activity (EDA) data.

Main Results:

• VIAStress demonstrated superior generalization performance on unseen subjects across four public datasets (WESAD, UBFC-Phys, VerBIO, CAN-STRESS).
• Achieved high accuracy in both within-dataset and cross-dataset, subject-independent settings.
• Outperformed existing competitive approaches in stress recognition tasks.

Conclusions:

• VIAStress offers a robust and adaptable solution for personalized stress recognition, accounting for user differences.
• The framework supports the development of more effective stress management strategies.
• Highlights the potential of instance-adaptive and multimodal domain generalization for personalized health monitoring.