Gaussian Process Regression for Predictive But Interpretable Machine Learning Models: An Example of Predicting Mental...

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Brain-computer interfaces (BCIs) are increasingly used for cognitive state monitoring.
• Machine learning-based BCIs often lack interpretability, functioning as 'black boxes'.
• Developing interpretable BCIs is crucial for understanding cognitive workload and improving human-computer interaction.

Purpose of the Study:

• To develop and evaluate an interpretable machine learning model for predicting cognitive workload.
• To compare the performance of Gaussian Process Regression (GPR) with Multiple Linear Regression (MLR) for workload prediction.
• To identify key electroencephalography (EEG) features relevant for cognitive workload assessment.

Main Methods:

• Participants performed N-back tasks with auditory-verbal, visual-spatial, and visual-numeric stimuli at varying cognitive loads.
• Gaussian Process Regression (GPR) and Multiple Linear Regression (MLR) models were trained and tested on EEG data.
• Feature importance analysis was conducted to identify predictive EEG features.

Main Results:

• The GPR model achieved an average standardized mean squared error (sMSE) of 0.44, outperforming the MLR model's sMSE of 0.55.
• A small subset of EEG features (top 25%) accounted for most of the GPR model's predictive power.
• GPR identified more efficient feature subsets compared to linear models like ANOVA.

Conclusions:

• Gaussian Process Regression (GPR) offers a powerful and interpretable approach for real-time cognitive workload monitoring using BCIs.
• Interpretable BCIs can achieve high predictive accuracy with fewer features, simplifying model development and deployment.
• This research paves the way for more transparent and efficient BCIs for cognitive state assessment.