Learning EEG Representations with Weighted Convolutional Siamese Network: A Large Multi-Session Post-Stroke Rehabilitation Study

Although brain-computer interface (BCI) shows promising prospects to help post-stroke patients recover their motor function, its decoding accuracy is still highly dependent on feature extraction methods. Most current feature extractors in BCI are classification-based methods, yet very few works from literature use metric learning based methods to learn representations for BCI. To circumvent this shortage, we propose a deep metric learning based method, Weighted Convolutional Siamese Network (WCSN) to learn representations from electroencephalogram (EEG) signal. This approach can enhance the decoding accuracy by learning a low dimensional embedding to extract distance-based representations from pair-wise EEG data. To enhance training efficiency and algorithm performance, a temporal-spectral distance weighted sampling method is proposed to select more informative input samples. In addition, an adaptive training strategy is adopted to address the session-to-session non-stationarity by progressively updating the subject-specific model. The proposed method is applied on both upper limb and lower limb neurorehabilitation datasets acquired from 33 stroke patients, with a total of 358 sessions. Results indicate that using k-Nearest Neighbor as the classification algorithm, the proposed method yielded 72.8% and 66.0% accuracies for the two datasets respectively, significantly better than the other state-of-the-arts ({p} < {0.05}). Without losing generality, we also evaluated the proposed method on two publicly available datasets acquired from healthy subjects, wherein the proposed algorithm demonstrated superior performance at most cases as well. Our results support, for the first time, the use of a metric learning based feature extractor to learn representations from non-stationary EEG signals for BCI-assisted post-stroke rehabilitation.