Improvement of EEG-based motor imagery classification using ring topology-based particle swarm optimization

Mental tasks classification such as motor imagery, based on EEG signals is an important problem in brain-computer interface (BCI) systems. One of the major concerns in BCI is to have an accurate classification. Classifier tuning is one of the most important techniques to increase classification accuracy. In this paper, a ring topology based particle swarm optimization (RTPSO) algorithm is proposed to tune classifiers. Fitness function of RTPSO algorithm is based on the 10-Fold Cross-Validation (CV) or Holdout methods which are used to evaluate performance of classifiers. Feed Forward Neural Network (FFNN) and three types of Support Vector Machine (SVM) classifiers are used to classify mental tasks. The proposed method tunes classifiers efficiently and quickly in a minimum of 10 iterations and outperforms the BCI 2003 and 2005 competition-winning methods and other similar studies on the same Graz datasets. Obtained results of the tuned FFNN proved far better than SVMs and classification algorithms of the other studies on the Graz datasets III and IIIb in all the experiments. According to the criterion of the BCI competition 2003 on the Graz dataset III, the maximal Mutual Information (MI) by tuned FFNN is about 0.81 while by the Least Squares SVM classifiers is about 0.73. FFNN improves misclassification rate comparing with the best of previous methods The mean of the maximal MI steepness is also improved. Our experiments show that the proposed RTPSO together with 10-Fold CV leads to promising results for classifier tuning in motor imagery classification.