An Algorithm for Motor Imagery Classification Based on Transfer Learning and Feature Fusion

Motor imagery-based brain computer interface (MI-BCI) using electroencephalogram (EEG) has attracted increasing attention due to its huge application potentials and low cost. However, decoding of MI-EEG signals is a challenging work because of low signal-to-noise ratio and high variability. This study aimed to develop an MI-EEG decoding algorithm with high performance. Specifically, we applied a transfer learning strategy to enhance transferability between EEG sessions. As an improvement of traditional common spatial pattern (CSP) algorithm, time-frequency common spatial patterns (TFCSP) were introduced to our method to extract narrowband information from time stages and frequency components of EEG signals. We fused narrowband information with broadband information extracted from CSP, selected informative features by Relieff algorithm. Finally, the optimized features were fed into the classifier to accomplish the classification and the performance of using multiple classifiers was compared. We verified the algorithm with a public dataset from BCI competition IV. The accuracy on test set reached to 89.20% and the cross-validation accuracy reached to 93.89 % when using support vector machine (SVM) as the classifier. Our approach and results suggest the huge potential of transfer learning and feature fusion strategy in MI-EEG decoding.