Multiclass self-paced motor imagery temporal features classification using least-square support vector machine

Mental tasks classification such as motor imagery based on EEG signals is a challenging issue in brain-computer interface (BCI) systems. Automatic classifier tuning seems to be an essential component in real-time BCI systems which makes the interface more reliable and easy to use and may offer the optimum configuration of classifier. This paper investigates the robustness of Least-Square Support Vector Machine (LS-SVM) to classify multi-class self-paced motor imagery (MI) temporal features while tuning the hyperparameters automatically. MI electroencephalogram (EEG) signals were preprocessed and segmented into non-overlapped distinctive time slots. Five different temporal features were extracted to characterize various properties of three Mis. An extended version of LS-SVM was employed for feature classification while the kernel model parameters were tuned by means of two optimization techniques, Coupled Simulated Annealing (CSA) followed by Simplex. LS-SVM parameters were evaluated and selected through leave-one-out cross validation (LOOCV) cost function. Finally, the proposed method was evaluated and compared to three widely used classifiers. The results indicated the high potential of LS-SVM to classify different Mis by obtaining the average classification accuracy 89.88±8.00 when using Sign Slop Changes (SSC) features. However, this LS-SVM performed slowly due to its additional steps for automatic model parameter tuning. In the comparative study, it was shown that each classifier behaved differently when various features were served; however, KNN outperformed others in both terms of classification accuracy and speed.