A Transductive Model-based Stress Recognition Method Using Peripheral Physiological Signals.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Current stress recognition methods struggle with individual physiological signal variations, leading to dispersed data and imbalanced samples.
• Global optimization classifiers in existing research do not adequately address the challenges of personalized stress recognition.

Purpose of the Study:

• To propose a novel framework for real-time stress recognition using peripheral physiological signals.
• To mitigate errors arising from individual differences in physiological data.
• To enhance the accuracy and regressive performance of stress recognition models.

Main Methods:

• A transductive learning framework utilizing epsilon-support vector regression (e-SVR) was developed.
• Non-linear real-time features were extracted via wavelet packet decomposition and bi-spectrum analysis.
• The model incorporates local learning principles, leveraging neighborhood knowledge and considering label dispersion.

Main Results:

• The proposed transductive model demonstrated superior prediction performance over traditional methods.
• Evaluation on the DEAP dataset and Stroop training confirmed the framework's effectiveness.
• Field studies validated the usability of the real-time interactive stress recognition framework.

Conclusions:

• The transductive learning approach offers a significant improvement for real-time stress recognition.
• The framework effectively handles individual differences in physiological signals.
• The developed method shows promise for practical applications in stress monitoring.