A Cross-Space CNN With Customized Characteristics for Motor Imagery EEG Classification

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) decode neurological activity for device control using motor imagery-electroencephalogram (MI-EEG) signals.
• Current MI-EEG classification methods face limitations in spatial resolution and subject-specific characterization, hindering accuracy, especially in multi-class tasks.

Purpose of the Study:

• To develop a novel cross-space convolutional neural network (CS-CNN) for enhanced four-class MI-EEG classification.
• To address limitations of single-space algorithms and insufficient subject specificity in existing BCIs.

Main Methods:

• Proposed a CS-CNN integrating modified customized band common spatial patterns (CBCSP) and duplex mean-shift clustering (DMSClustering) for cross-space feature extraction.
• Extracted multi-view features from time, frequency, and space domains, fusing them within the CNN for classification.
• Utilized MI-EEG data from 20 subjects, including real MRI information for some analyses.

Main Results:

• Achieved classification accuracy of 96.05% with MRI data and 94.79% without MRI on a private dataset.
• Demonstrated superior performance on the BCI Competition IV-2a dataset, outperforming state-of-the-art algorithms with a 1.98% accuracy improvement and a 5.15% reduction in standard deviation.

Conclusions:

• The proposed CS-CNN effectively integrates cross-space information and subject-specific characteristics for robust MI-EEG classification.
• This approach significantly enhances BCI performance, offering a promising direction for advanced neurological decoding and control applications.