Self-organized phase-locked component pair learning in motor imagery classification

Phase synchronization measurements among electroencephalogram (EEG) channels and single-channel power spectrum characterize brain activities during motor imagery (MI) from different angles. Recently, end-to-end deep learning (DL) networks have achieved remarkable success by simultaneously optimizing preprocessing, power spectrum feature extraction, and classification procedures. However, no end-to-end network has utilized both phase synchrony measurement and power spectrum, which mutually contribute valuable and complementary information for MI classification. Extracting interpretable phase synchrony-based features such as phase locking value (PLV) through DL architecture is challenging due to the computational complexity. To tackle this challenge, we propose a transplantable convolution module called phase-to-amplitude transcoder to transform phase differences between pairs of signals into specialized amplitude representations, whose data distribution is sensitive to MI types. Preprocessing and pairing modules are designed to provide high-SNR signal pairs for this transcoder. Another convolution module is used for classification based on the data distribution. By integrating these modules, we constructed the first end-to-end classification network based on phase-locking information. Extensive experiments have demonstrated that our network outperforms state-of-the-art methods in classifying 1s-long MI samples. Signal pairs learned by the trained network exhibit high PLVs, confirming the transcoder's effectiveness and the network's capability to self-organize phase-locked component pairs (PLP). Notably, a certain signal pair extracted from the tongue MI samples of a subject is significantly phase-locked, with half of them exhibiting PLVs over 0.937. This high level of phase-locking has not been found in existing literature. Using DL to explore PLPs to investigate phase synchronization represents a promising research direction.