Performance Comparison of Different Classifiers to Detect Motor Intention in EEG-Based BCI

In this work, the performance of different types of classifiers to detect event-related desynchronization of EEG sensorimotor rhythms in brain-computer interfaces was evaluated. The study used an 8 healthy volunteers EEG signals database, it was developed at the Neuromuscular and Sensory Research and Rehabilitation Engineering Center at the Faculty of Engineering of the National University of Entre Ríos. The feature extraction method was the power spectral density (PSD) and four types of classifiers were used: lazy (K-nearest neighbors), connectionist (multilayer perceptron), kernel-based (support vector machine) and statistical (linear discriminant). Although Accuracy (Acc) and True Positive Rate (TPR) were reported for the classifiers performance, the last one was taken as a more appropriate metric due to the BCI application in rehabilitation. The results obtained for classifier were Lazy (Acc = 60% and TPR = 69%), Connectionist (Acc = 68%, TPR = 46%), kernel-based (Acc = 63%, TPR = 88%), statistical (Acc = 59%, TPR = 74%), the results show that the highest TPR belongs to less complex classifiers, such as linear support vector machine and linear discriminant.