Improved Motor Imagery Classification Using Regularized Common Spatial Pattern with Majority Voting Strategy

The classification of motor imagery (MI) can be substantially improved when the electroencephalogram (EEG) features are extracted using the Common Spatial Pattern (CSP) algorithm. However, most of the previous studies have empirically selected a window size of Is to extract the CSP-features and did not employ any post-processing technique such as the majority voting technique. The aim of this study is to classify hand and foot movement tasks using EEG data from fourteen healthy subjects (20-30 years) and four machine-learning (ML) algorithms such as Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), k Nearest Neighbor (KNN), and Random Forest (RF). The CSP features were extracted from various window sizes ranging between 0.3s to 2s. In the post-processing stage, the number of votes was varied between 1 and 9. The results show that the KNN can achieve 78.9% accuracy using 2s window sizes with 9 votes. However, the ranking analysis using both accuracy and area under the curve values reveals that the RF algorithm can consistently perform well using different window sizes and number of votes. The prediction accuracy was higher for foot MI compared to hand MI; however, the difference was not significant (p>0.05). The overall mean accuracy could be improved by 13.7% using LDA, 8.2% using SVM, 13.7% using KNN, and 6.6% using RF while varying window size and number of votes. The results of this study are valuable for improving MI task classification as well as in developing control function for prosthetics and exoskeleton devices. Copyright (C) 2021 The Authors.