Stepwise Discriminant Analysis based Optimal Frequency Band Selection and Ensemble Learning for Same Limb MI Recognition

Abstract

Same limb motor imagery (MI) brain-computer interfaces can effectively overcome the cognitive disassociation problem of the traditional different-limb MI paradigm, and they can reduce the patient burden and extend the functionality of external devices more effectively. However, the electroencephalogram (EEG) MI features of same limb originate from one side of the brain, which poses a great challenge to MI EEG feature mining and selection as well as accurate decoding. To overcome this problem, we propose an adaptive feature selection strategy for subject-specific optimal frequency band based on regularized common spatial pattern (RCSP) and stepwise discriminant analysis, then combine the integrated classification strategy to accurately decode three types of single-limb MI tasks. As there are minor frequency band differences and huge variability for the same limb MI tasks, the optimal frequency band range for each subject was selected by stepwise discriminant analysis, and RCSP was used to extract spatial distribution features, which reduced the influence of the length of the time window and differences of the frequency bands. Then an integrated classification strategy based on multiple efficient classifiers is used for MI accurate recognition. The proposed method obtains 76.58% accuracy in the unilateral limb MI recognition task, which is 12.67%, 9.89%, 6.62%, and 7.90% higher than other traditional decoding methods such as CSP + LDA, FBCSP + LDA, FBCSP + C2CM, and FBCSP + SVM, respectively. Compared with Deep ConvNet and EEGNet, the decoding accuracy is improved by 16.93% and 7.33%, respectively. The experimental results show that our proposed highly efficient method improves the decoding accuracy for classifying different joints of unilateral limbs and has high promotion and application value.