Zero-Calibration MI Decoding via Self-Supervised Representation and Ensemble Learning.

Abstract

Motor Imagery (MI) is a core task in brain computer interface (BCI) research. Zero-calibration MI de coding, which eliminates the need for training subject specific models, significantly reduces data annotation and training time, thus attracting widespread attention. However, this paradigm demands a higher level of precision in extracting invariant features. This study proposes an efficient strategy that combines self-supervised representation learning with supervised classification learning to accurately extract intrinsic invariant features from EEG signals. Specifically, through random masking and feature reconstruction mechanisms, the encoder performs self-supervised learning to uncover universal EEG signal features. Additionally, an ensemble learning classifier is used to further compress features and significantly enhance model performance through multi-branch comprehensive decision-making. Our method employs a pure convolutional neural network (CNN) architecture, achieves excel lent performance in three MI tasks: in the complex four class classification task, it is the only method to achieve an accuracy exceeding 60%; in the two binary classification tasks, accuracies of 85.50% and 82.98% were achieved. In the comparison, the p-values of almost all methods were less than 0.05, demonstrating significant statistical significance. This study provides an innovative and efficient solution for cross-subject zero-calibration MI decoding.