Predicting cognitive load with EEG using Riemannian geometry-based features.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Cognitive load (CL) estimation is crucial for adaptive systems.
• Existing electroencephalography (EEG)-based CL prediction models face challenges in generalizability.
• Riemannian geometry offers a novel framework for analyzing complex signal data.

Purpose of the Study:

• To enhance cognitive load (CL) prediction accuracy using electroencephalography (EEG) data.
• To investigate the efficacy of Riemannian geometry features for CL prediction.
• To evaluate the generalizability of models across unseen subjects.

Main Methods:

• Utilized Riemannian geometry features, including spatial covariance and correlation matrices of EEG signals and their derivatives.
• Employed Riemannian Procrustes Analysis (RPA) for feature extraction and model generalization.
• Evaluated performance using the Minimum Distance to Riemannian Mean model and compared against baseline methods.

Main Results:

• The spatial covariance matrix of the EEG signal's first-order derivative significantly improved prediction performance.
• Riemannian Procrustes Analysis (RPA) demonstrated superior generalizability with limited calibration data.
• The proposed approach outperformed existing methods for cognitive load prediction.

Conclusions:

• Riemannian geometry, particularly RPA, offers a promising avenue for robust, generalizable cognitive load prediction from EEG.
• Novel features derived from signal derivatives enhance performance in Riemannian-based analyses.
• This approach holds significant potential for real-world applications requiring efficient, cross-subject CL estimation.