Learning EEG Representations With Weighted Convolutional Siamese Network: A Large Multi-Session Post-Stroke...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) offer potential for post-stroke motor function recovery.
• Current BCI decoding accuracy is limited by feature extraction methods, predominantly classification-based.
• Metric learning methods are underutilized for learning BCI representations.

Purpose of the Study:

• To propose a novel deep metric learning method, Weighted Convolutional Siamese Network (WCSN), for improved electroencephalogram (EEG) signal representation.
• To enhance BCI decoding accuracy for post-stroke neurorehabilitation.
• To address challenges of training efficiency and non-stationarity in EEG data.

Main Methods:

• Developed a Weighted Convolutional Siamese Network (WCSN) for deep metric learning on EEG data.
• Implemented a temporal-spectral distance weighted sampling method for informative sample selection.
• Utilized an adaptive training strategy to manage session-to-session non-stationarity.

Main Results:

• The WCSN method achieved 72.8% and 66.0% accuracies on upper and lower limb neurorehabilitation datasets, respectively, outperforming state-of-the-art methods.
• Superior performance was also observed on publicly available datasets from healthy subjects.
• Demonstrated the efficacy of metric learning for feature extraction from non-stationary EEG signals.

Conclusions:

• The proposed WCSN method significantly improves BCI decoding accuracy for post-stroke rehabilitation.
• Metric learning-based feature extraction is effective for non-stationary EEG signals in BCI applications.
• This study provides foundational evidence for using metric learning in BCI-assisted neurorehabilitation.