Decoding motor imagery related to major mimetic muscles from electroencephalography

Background Functional and aesthetic deficits in individuals with facial nerve paralysis (FNP) significantly impair their quality of life. By decoding motor intentions and controlling rehabilitation devices, motor imagery (MI)-based brain-computer interfaces can improve outcomes in people with peripheral paralysis. However, the electroencephalography (EEG) features underlying different facial MIs and their decodability remain unclear. This study aims to investigate the feasibility of achieving accurate decoding of facial MIs related to major mimetic muscles and the corresponding decoding strategies. Methods After comparing block and event-related designs to identify the appropriate paradigm for facial MIs, 20 healthy participants performed four types of facial MIs (eyebrow raising, eye closing, lip puckering and grinning) in two modalities: kinesthetic and visual, from which event-related desynchronization/synchronization (ERD/S) features were extracted using time-frequency analysis. A deep learning model integrating a temporal convolutional network with a spatial attention mechanism was then developed for both within-subject and cross-subject decoding, thereby identifying the contribution of each EEG channel. Finally, the model was further evaluated on EEG data from six individuals with FNP. Results Participants showed better performance in the block design, in which facial MIs induced significant ERD in the low-frequency band in the left prefrontal and right central-temporal regions, co-occurring with shorter and weaker ERS in higher frequencies. Regarding MI decoding in healthy participants, the model achieved the highest average accuracy of 85.17% in within-subject classification of kinesthetic MI, with EEG features from the left frontal and parietal regions contributing most to decoding. Combining these findings, the model obtained an average accuracy of 76.46% on patients’ data, with half the number of MI tasks and 25% fewer EEG channels. Conclusion This study demonstrated that major mimetic muscle-related MIs can be accurately recognized from EEG using deep learning, with a suitable decoding strategy involving within-subject decoding of kinesthetic MI collected through a block design.