Features Extraction Method for Brain-Machine Communication Based on the Empirical Mode Decomposition

Abstract

A brain-machine interface (BMI) is a communication system that translates human brain activity into commands, and then these commands are conveyed to a machine or a computer. It is proposes a technique for features extraction from electroencephalographic (EEG) signals and afterward, their classification on different mental tasks. The empirical mode decomposition (EMD) is a method capable of processing non-stationary and nonlinear signals, as the EEG. The EMD was applied on EEG signals of seven subjects performing five mental tasks. Six features were computed, namely, root mean square (RMS), variance, Shannon entropy, Lempel–Ziv complexity value, and central and maximum frequencies. In order to reduce the dimensionality of the feature vector, the Wilks' lambda (WL) parameter was used for the selection of the most important variables. The classification of mental tasks was performed using linear discriminant analysis (LDA) and neural networks (NN). Using this method, the average classification over all subjects in database is 91 ± 5% and 87 ± 5% using LDA and NN, respectively. Bit rate was ranging from 0.24 bits/trial up to 0.84 bits/trial. The proposed method allows achieving higher performances in the classification of mental tasks than other traditional methods using the same database. This represents an improvement in the brain-machine communication system.

An optimized feature selection and classification method for using electroencephalographic coherence in brain–computer interfaces

“…As in the results of the data in Section 3.1, here we note that the classification achieves a good performance and with the advantage of using only three sensors blindly selected. Furthermore, our results can be also compared against those recently reported in [19] for the same data sets. There, the authors report an average accuracy in the classification of the same two-classes combinations of up to 98.7% for Subj7 (best performing subject).…”

“…Such process ends up generating (6 EEG channels) × (5 mode functions) × (6 metrics) = 180 elements in the feature vector. Hence, in [19] the authors propose a secondary feature selection method by which they are capable to reduce the dimension of their feature vector to 16 ± 7 (depending on the subject and on the mental task), but with a reduction in the average performance to 90.8% for Subj7. In our case, the overall performance for Subj7 was 89.1%, but using a three-dimensional feature vector and with the capability of gaining insight of the brain process through the selected sensors and their connectivity.…”

“…A final test performed on this data set corresponded to evaluate the classification accuracy according to the settings in [19], whose result in terms of the mean and standard deviation among all possible tasks combinations for each subject is shown in Fig. 3.…”

“…Error bars show a ± standard deviation. The bars labeled as COH correspond to the results of the method here proposed, whileLD, 20-10NN, and 10-5NNcorrespond to the results using the feature vectors and different classification strategies proposed in[19].…”

Black hole algorithm with convolutional neural networks for the creation of brain-computer interface based in visual perception and visual imagery

“…In order to evaluate the different options, it has been chosen to use the accuracy and Cohen's Kappa value, widely used in the literature [29,30]. Together, they give us an explanation of how the classification model behaves:…”