Gaussian Processes with Physiologically-Inspired Priors for Physical Arousal Recognition.

Area of Science:

• Physiological computing
• Machine learning
• Biomedical signal processing

Background:

• Machine learning models struggle with small or noisy datasets, potentially missing crucial domain-specific information.
• Expert knowledge is vital but difficult to integrate directly into standard algorithms.
• Gaussian Process Classification (GPC) offers a Bayesian framework suitable for incorporating prior knowledge.

Purpose of the Study:

• To investigate the effectiveness of Gaussian Process Classification (GPC) for integrating domain-specific knowledge into machine learning.
• To enhance the training phase of algorithms by injecting problem-specific information.
• To improve the recognition of physical stressors using physiological data.

Main Methods:

• Implemented domain knowledge by defining an a-priori distribution on the GPC latent function.
• Extracted handcrafted features from Heart Rate Variability (HRV) and Electrodermal Activity (EDA) signals.
• Correlated physiological features to classification logits using a physiology-informed prior function.
• Updated the GPC model using Bayes' formula with subject data.

Main Results:

• Comparative experiments demonstrated the selection of effective physiologically-inspired GPC prior functions.
• The recognition accuracy of a physical stressor was significantly enhanced.
• Physiologically-informed prior knowledge injection improved GPC model performance.

Conclusions:

• Gaussian Process Classification (GPC) is a suitable model for implementing domain knowledge in machine learning.
• Integrating physiological prior knowledge into GPC models improves the analysis of physiological signals.
• This method offers a robust approach for stressor detection in autonomic nervous system dynamics.