Deep Channel-Correlation Network for Motor Imagery Decoding From the Same Limb.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) is a key brain-computer interface (BCI) paradigm enabling control without external stimuli.
• Decoding MI of different joint movements from the same limb offers intuitive device control but faces limited research and accuracy.
• Current limitations hinder the practical application of MI-based BCIs.

Purpose of the Study:

• To explore the decoding performance ceiling for three tasks: resting state, right-hand MI, and right-elbow MI using deep learning.
• To develop and evaluate a novel deep learning approach for decoding fine-grained motor imagery.
• To assess the feasibility of practical applications for MI of different joints from the same limb.

Main Methods:

• Utilized correlation matrices derived from electrode data to represent functional brain relationships.
• Proposed a Channel-Correlation Network (CCN) for learning channel representations for classification.
• Applied ensemble learning to integrate multiple CCNs for improved decoding accuracy.

Main Results:

• Achieved a decoding accuracy of up to 87.03% in a 3-class scenario (resting state, right-hand MI, right-elbow MI).
• Demonstrated the effectiveness of the proposed deep learning method in distinguishing between different joint movements.
• Validated the potential of the fine-grained MI paradigm for practical BCI applications.

Conclusions:

• Deep learning methods are effective for decoding motor imagery of different joints within the same limb.
• The Channel-Correlation Network combined with ensemble learning shows significant promise for advancing BCI technology.
• The findings support the practical viability of this fine-grained motor imagery paradigm.