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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) based brain-computer interfaces (BCIs) show potential for stroke neurorehabilitation.
• Current MI-BCI systems face challenges with inter-subject variability, limited training data, and lengthy calibration periods.

Purpose of the Study:

• To develop a novel framework for calibration-free MI-EEG decoding.
• To address limitations of existing MI-BCI systems, including inter-subject variability and the need for extensive training data.

Main Methods:

• A Meta-Learning Enhanced Multi-Source Domain Adaptation (MLEMSDA) framework was proposed, unifying cross-task, cross-dataset, and cross-subject domain adaptation.
• Gradient-based meta-learning was employed for calibration-free decoding, utilizing pre-training on public datasets and meta-learning fine-tuning on target datasets.
• The framework was tested on unseen subjects using a leave-one-out cross-validation approach.

Main Results:

• The MLEMSDA framework achieved high classification accuracies on multiple MI-EEG datasets: 77.87% (DeepConvNet on CBCIC), 75.54% (EEGNet on own dataset), and 72.72% (ShallowConvNet on BCI Competition IV dataset 2b).
• The proposed method demonstrated superior classification accuracy in a zero-calibration scenario compared to competing methods.
• Validation was performed on public and collected MI-EEG datasets, including stroke patients.

Conclusions:

• The MLEMSDA framework effectively enables calibration-free MI-EEG decoding.
• The method shows strong generalizability and effectiveness, paving the way for more practical MI-BCI applications in neurorehabilitation.