Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

Patients that have suffered a stroke or traumatic brain injuries can have their motion capability reduced. The use of motion assistant devices controlled by a brain-machine interface (BMI) can improve the rehabilitation process through the cognitive involvement of the subject [1].

One of the most used BMI control approach involves the use of electroencephalographic (EEG) signals due to its non-invasive nature. A BMI collects the EEG signals of the subject’s brain through several electrodes. The channel information is then analyzed by a computer in order to extract the representative features of the waves. Once the classification model is created, it can be used to identify similar patterns in the real-time performance, controlling the assistant device depending on the identified mental will.

Motor imagery (MI) is one of the common used control methods. It consists of the mental imagination of the movement action without real movement. Literature indicates that there is an event related (de)synchronization (ERD/ERS) [2] mainly at the mu frequency band (10–14 Hz), as a part of alpha band (8–13 Hz), consisting of a fluctuation of the power due to the MI actions in comparison to the relaxed state. This is more noticed on the zone around the CZ electrode based on the 10/10 international system.

In this paper, besides the previously mentioned alpha band (8–13 Hz), beta (13–32 Hz) and gamma bands (32–50 Hz) [3] were also considered in order to assess the attention focus [4] of the subject, and to improve the results of the MI identification. In addition, three different processing algorithms and classifiers were studied.

2 Materials and Methods

2.1 Subjects

Three healthy subjects (1 male and 2 women) participated voluntary in the experimental sessions. The users were previously informed about the procedure and signed an informed consent according to the Helsinki declaration. The whole experimental procedure was approved by the ethics committee of the University.

2.2 Equipment

EEG acquisition was performed by a Starstim 32 cap of Neuroelectrics™. The sampling frequency was 500 Hz and the data were transmitted by wire to the computer where they were analyzed with the help of the custom algorithms developed in Matlab™. Although, all the electrodes were used for spatial filtering, only the electrodes around the sensory-motor cortex zone were considered for the data characterization: CZ, CP1, CP2, C1, C2, C3, C4, FC1 and FC2.

2.3 Experimental Setup

Every subject performed 5 trials. Each of them consisted of 15 consecutive events of relax/imagination, while the subject was comfortably seated. Each event lasted around 5 s and required the subject not to move or to blink. During the event, the subjects had to focus on the mental action of pedaling during imagination periods or to leave the mind in a blanked state during the relaxed periods. Between events, a little time of 2 s was allowed to the subject in order to blink or to move during the trial. Event information was communicated to the subject by a screen interface during the whole trial.

2.4 Signal Processing

EEG signals were processed in 1 s epochs shifted every 0.2 s. Data were preprocessed by a high pass filter (0.05 Hz), a low pass filter (45 Hz) and a Laplacian spatial filter.

For the pattern characterization, three methods were considered to extract one feature per alpha, beta and gamma bands:

  • ST: Value of the peak of the Hilbert Marginal Spectrum [5], obtaining the analytical signal with the Stockwell transform [6]

  • PSD: Value of the peak of the power spectral density by Welch’s method

  • H: Value of the amplitude of the most representative harmonics per band obtained by Fast Fourier Transform.

Accuracy evaluation was done by leave-one-out cross-validation. Three different classifiers were also considered: Support Vector Machine (SVM) [7], Linear Discriminator Analysis (LDA) [8] and Self-Organizing Maps (SOM) [9]. Accuracy represents in percentage the number of events correctly detected.

3 Results

Table 1 shows the accuracy performance of the classifiers. While SVM and LDA present a similar result, SOM shows a lower accuracy. Table 2 shows the results by subject, classifier and type of event neglecting SOM classifier, due to its lower performance, to limit the volume of data.

Table 1. Accuracy performance of the classifiers.
Full size table
Table 2. Accuracy results for the 3 subjects, best 2 classifiers and relax and imagine events.
Full size table

Statistical Manova analysis by SPSS revealed that accuracy had a high dependency on the subject’s expertise (p < 0.001). The test of within-subjects effects indicated that there was not a dependency on the subjects regarding the methods (p > 0.05), or the classifier (p > 0.05). However, the performance of the type of event (relax/imagination) showed a clear dependency on the subject (p < 0.001).

4 Discussion

Results indicated that SOM performance was lower than SVM or LDA classifiers. Time of computing was also higher. The SOM performance could be improved by an algorithm tuning and number of trainings. However, the simplicity and faster computing of SVM or LDA makes them more suited for a BMI. BMI performance is very dependent on the subject’s expertise as S1 results indicate. ST obtained the best results for all the subjects with a better result of SVM during relax and LDA during imagine. However, as in an online application, the type of event is not known at first hand, it is not possible to use both.

5 Conclusion

The paper has introduced several MI algorithms and classifiers. A combination of ST+SVM or LDA would be the best for a future real-time application.