Physiologically-Informed Gaussian Processes for Interpretable Modelling of Psycho-Physiological States.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Machine Learning

Background:

• Machine Learning (ML) models are increasingly used in healthcare, raising concerns about interpretability and accountability.
• Interpretable ML is crucial for trust and validation in clinical applications.

Purpose of the Study:

• To introduce the Physiologically-Informed Gaussian Process (PhGP) classification model.
• To enhance the interpretability and accuracy of ML models in healthcare by incorporating physiological domain knowledge.

Main Methods:

• Developed PhGP, an interpretable ML model based on Gaussian Processes (GPs).
• Injected physiological domain knowledge as a prior distribution into the GP latent space.
• Utilized raw physiological signals and a prior function for hyper-parameter estimation.
• Introduced novel interpretability metrics to identify informative input regions.

Main Results:

• PhGP demonstrated improved classification performance compared to competitive methods across three datasets.
• The model successfully provided physiologically sound interpretations for its predictions.
• Evaluated on electrodermal activity (EDA) signals during emotional, painful, and stressful tasks.

Conclusions:

• PhGP offers a promising approach for accurate and interpretable ML in healthcare.
• Incorporating physiological priors enhances model performance and provides meaningful insights.
• The developed interpretability metrics aid in understanding model predictions in physiological contexts.