Meta-Learning Enhanced Multi-Source Domain Adaptation for zero-calibration motor imagery EEG decoding.

BACKGROUND: Motor imagery (MI) based brain-computer interface (BCI) holds promising application prospects for closed-loop neurorehabilitation in stroke recovery. Despite substantial progress, challenges such as inter-subject variability, lack of training data for specific subject, and the need for time-consuming calibration still hinder the practical deployment of MI-BCI systems. NEW METHOD: In this work, aiming to address these issues, we propose a novel Meta-Learning Enhanced Multi-Source Domain Adaptation (MLEMSDA) framework that unifies cross-task, cross-dataset, and cross-subject domain adaptation with gradient-based meta-learning to enable calibration-free MI-EEG decoding. Specifically, two large public ME and MI EEG datasets are firstly used for pre-training to facilitate cross-task and cross-dataset knowledge transfer. Afterward, to further reduce the differences in feature distribution among different individuals, meta-learning based fine-tuning is performed using data from all subjects in the target dataset except the unseen subject. Finally, the obtained decoding model is tested on the unseen subject. RESULTS: The proposed MLEMSDA framework was validated on a public stroke MI EEG dataset (CBCIC), our own collected MI EEG dataset, and BCI Competition IV dataset 2b using leave-one-out cross-validation method. DeepConvNet achieved the highest average accuracy of 77.87% on CBCIC dataset, EEGNet yielded the highest average accuracy of 75.54% on our own collected dataset, and ShallowConvNet obtained the highest average accuracy of 72.72% on BCI Competition IV dataset 2b. COMPARISON WITH EXISTING METHODS: With respect to classification accuracy in the zero-calibration scenario, our method outperforms all the competing methods. CONCLUSION: These results clearly demonstrate the effectiveness and generalizability of our method, paving the way for more practical MI-BCI applications.