Periodicity-based multi-dimensional interaction convolution network with multi-scale feature fusion for motor imagery EEG classification

Abstract

Background

The Motor Imagery (MI)-based Brain-Computer Interface (BCI) has vast potential in fields such as medical rehabilitation and control engineering. In recent years, MI decoding methods based on deep learning have gained extensive attention. However, capturing the complex dynamic changes in EEG signals remains a challenge, and the decoding performance still needs further improvement.

New methods

The paper proposes a novel method, Periodicity-based Multi-Dimensional Interaction Convolution Network with Multi-Scale Feature Fusion (PMD-MSNet), for MI-EEG signal classification. It converts 1D EEG signals into multi-period 2D tensors to capture intra-period and inter-period variations and enables cross-dimensional interaction based on periodic features. Subsequently, parallel multi-scale convolution is utilized to adaptively extract temporal, frequency, and time-frequency features.

Results

Experimental results on the BCI IV-2a dataset demonstrate that the PMD-MSNet model achieves a classification accuracy of 82.25 % on average and a kappa value of 0.763, which significantly outperforms seven other deep learning-based EEG decoding models. The model attained the highest classification accuracy and kappa value among the seven subjects, showcasing its superior performance and robustness.

Conclusions

The PMD-MSNet model incorporates periodic features, multi-dimensional interaction mechanisms, multi-scale convolutions to achieve efficient feature extraction and classification of EEG signals, significantly enhancing the performance of MI classification tasks.