Motor imagery decoding network with multisubject dynamic transfer.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) offer intelligent rehabilitation for motor function.
• Accurate decoding of motor imagery electroencephalogram (MI-EEG) signals is crucial for BCI effectiveness.
• Inter-individual differences in EEG signals necessitate dynamic adaptation in decoding models.

Purpose of the Study:

• To propose a novel multi-source dynamic conditional domain adaptation network (MSDCDA) for enhanced MI-EEG decoding.
• To address the challenge of multi-source domain conflicts and negative transfer in existing domain adaptation methods.
• To improve the generalization and decoding performance of BCI models across different subjects.

Main Methods:

• Utilized a multi-channel attention block to focus on relevant EEG channels.
• Employed a spatial-temporal convolution block for shallow feature extraction.
• Incorporated a dynamic residual block for subject-specific feature adaptation.
• Applied Margin Disparity Discrepancy (MDD) with adversarial learning for conditional domain adaptation.

Main Results:

• Achieved decoding accuracies of 78.55% on Dataset IIa and 85.08% on Dataset IIb of BCI Competition IV.
• Demonstrated effective mitigation of multi-source domain conflicts.
• Significantly enhanced decoding performance for target subjects.

Conclusions:

• The proposed MSDCDA network effectively addresses multi-source domain conflicts in MI-EEG decoding.
• MSDCDA significantly improves BCI performance for motor function rehabilitation.
• This approach facilitates the broader application of BCI technology in clinical settings.