A novel ANN adaptive Riemannian-based kernel classification for motor imagery

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Riemannian geometry offers a novel approach to classifying electroencephalographic (EEG) covariance matrices.
• Existing methods often utilize raw EEG data, but classifying covariance matrices directly shows promise.
• Brain-computer interfaces (BCIs) rely on accurate signal classification for effective control.

Purpose of the Study:

• To introduce a new Artificial Neural Network (ANN) approach, the Adaptive Riemannian Kernel ANN (ARK-ANN), for EEG motor imagery signal classification.
• To evaluate the performance of ARK-ANN against traditional methods like Support Vector Machines (SVM).
• To assess the impact of a geodesic filter on classification accuracy within the ARK-ANN and ARK-SVM frameworks.

Main Methods:

• Utilizing covariance matrices derived from EEG signals instead of raw data.
• Implementing a multilayer perceptron with an adaptive optimization for the testing set.
• Applying Riemannian geometry principles for direct classification of covariance matrices.
• Incorporating a geodesic filter to enhance classification performance.

Main Results:

• The ARK-ANN achieved higher accuracy in inter-subject classification (87.4%) compared to the reference ARK-SVM (86%).
• For cross-subject classification, ARK-ANN reached 77.3% accuracy, improving precision by 8.2% over SVM-based methods.
• The geodesic filter provided marginal improvements in the ARK-SVM for inter-subject classification.

Conclusions:

• The proposed ARK-ANN method offers a more effective approach for classifying EEG motor imagery signals in BCI applications.
• Riemannian geometry combined with adaptive kernels presents a powerful strategy for EEG data analysis.
• Further research can explore geodesic filters to optimize Riemannian-based classification techniques.