THE use of motor imagery as A WAY to communicate between a brain and a computer

F. Cincotti¹, F. Babiloni², C. Babiloni², F. Carducci², J. Millán³, S. Salinari⁴, L. Bianchi⁵, M. G. Marciani^1,5
1 Fondazione Santa Lucia - IRCCS, Rome, Italy.    ² Dip. Fisiologia umana e Farmacologia, Università "La Sapienza", Rome, Italy.    ³ ISIS, Joint Research Center of the EC, Ispra (VA), Italy.    ⁴ Dip. di Informatica e Sistemistica, Università "La Sapienza", Rome, Italy.    ⁵ Dip. di Neuroscienze, Università "Tor Vergata", Rome, Italy.

Abstract: The opening of a communication channel between brain and computer (Brain Computer Interface; BCI) is possible by using EEG changes in power spectra related to the imagination of movements. In this paper we present results obtained by recording mental imagery of upper limbs in a total of 18 subjects by using low resolution SL, different linear, quadratic and non linear classifiers as well as a variable number of scalp electrodes, from 2 to 26. The results presented (variable recognition rate of mental imagery between 75 to 95%) suggest that it is possible to recognize mental imagery with a different kind of computational schemes for the generation of a BCI.

INTRODUCTION

The opening of a communication channel between brain and computer (Brain Computer Interface; BCI) is possible by using EEG changes in power spectra related to the imagination of movements. Such EEG variations are specifically located in centro-parietal scalp areas and can be recognized by on-line classifiers. Scalp potential distribution of EEG power spectra can be also enhanced spatially by using the Surface Laplacian (SL) operator. However, the issue of using as few scalp electrodes as possible for practical reasons pushed to use low-resolution SL operators in BCI field. Enhanced EEG power spectra distributions related to mental imagery can be recognized by using linear and non linear classifiers. In this paper we presented results obtained by recording mental imagery of upper limbs in a total of 18 subjects by using linear or quadratic classifiers as well as a variable number of scalp electrodes, from 2 to 128.  The results presented (variable recognition rate of mental imagery between 75 to 95%) suggested that it is possible to affordable recognize mental imagery with a different kind of computational schemes for the generation of a BCI.

METHODS

EEG recordings

The EEG potentials are recorded from a group of 18 subjects with an extension of the 10/20 International system (28 electrodes). Depending on the particular study described below, different subsets of such scalp electrodes were used to select the features to be classified.

The EEG sampling rate is 128 Hz. The main operation in the temporal domain is a spatial filtering whereby new potentials should represent better the cortical activity only due to local sources below the electrodes. We use the Welch periodogram algorithm to estimate the power spectrum of each signal. Windows are 0.5 seconds long, what gives a frequency resolution of 2 Hz. The values in the frequency band 8-30 Hz are normalized according to the total energy in that band. Thus, an EEG sample is represented by n*12 features, where 12 are the spectral component for each one of the n channels used. The periodogram, and hence an EEG sample, is computed every 1/2 second.

Recognition of Motor Imagery Tasks with Two Channels

A key issue in the developing of an efficient BCI device is that related to the use of a minimal number of electrodes. In order to explore the possibilities to classify EEG signals by using a reduced set of electrodes, we performed a study involving a large umber of normal subjects recorded while they are imaging right or left hand movements. The classifier was a implementation of the concept of the Mahalanobis distance. This classifier has only one prototype per mental task, computed as the mean vector of that class estimated using the corresponding EEG samples in the training set. The response of the classifier for the current EEG sample is just the class with the nearest prototype based on the Mahalanobis distance, which can be computed using either the full or diagonal covariance matrix for each mental task. EEG was gathered from a group of 13 healthy subjects performing two motor-related mental tasks, namely the imagination of right and left middle finger hand movements. Electrodes were placed in scalp centro parietal zones (C3, P3, C4 and P4 positions of the international 10-20 system). In this case, we extracted the power spectral density features as described before directly from the raw potentials without any SL transformation. In a few subjects for whom it was available a realistic head model based on sequential MR images, we estimated the cortical activity during the mental imagery task using a linear estimation algorithm. Such reconstruction was performed in order to check if the employed centro-parietal scalp electrodes (C3, P3, C4, P4) were appropriately located to gather the relevant electrical patterns elicited by the subject's mental activity from the cortical surface.

Figure 1. Estimated cortical current distribution of the power of EEG oscillations in the frequency range of 9-11 Hz for the subject FB during the mental imagery of right hand movements. The gray scale coded from white to light gray the desynchronization of the estimated current EEG rhythms with respect to the rest period, while from light gray to dark gray is coded the synchronization.

RESULTS

Figure 1 shows the decrease (from red to white) or the increase (from light to dark blue) of the power of EEG oscillations during the mental imagery of the right middle finger movements with respect to the power estimated in the rest period in the frequency band of 9-11 Hz for the subject F.B. This figure indicates that the induced decrement of the EEG oscillations is mainly focused in the cortical zone belonging to the primary motor area controlateral to the imagined movements. More important in this context, as illustrated in the figure, is the fact that such cortical area appears properly sampled by the C3 electrode, while the P3 electrode is positioned at the border of the desynchronized active cortical zone. To test the recognition capabilities of the classifiers, we have employed the k-fold cross-validation procedure for each subject, with k=10. Then, we did a statistical analysis of the average recognition scores. In particular we performed a two-way analysis of variance (ANOVA), using the classifiers (Mahalanobis-based with full, MAF, or diagonal, MAD, matrix and SSP) and the number of electrodes (2 or 4) as main factors. 

The main result is that the MAD classifiers were able to detect EEG activity related to imagination of movement with an affordable accuracy (average recognition scores 97% or higher) by using only the C3 and C4 electrodes. The ANOVA test confirms that the increase in performance of the two MAF classifiers is statistically significant with respect to any other with a p < 10^-5. It also shows that the slight improvement of the recognition rates obtained by the MAF with 4 electrodes with respect to that using just 2 electrodes (99.3% vs. 97.3%) is not statistically significant (p < 0.90).

DISCUSSION

In this paper we described some of the computational mechanisms useful to detect the imagination of motor acts in normal subjects by using EEG. We observed that the best features for the classification of such mental patterns are the power spectra decrease/increase with respect to a baseline period. Results obtained suggest that it is possible to recognize mental imagery with only two electrodes by using a quadratic classifier based on the computation of Mahalanobis distance. These results are in line with recent advancements in the BCI area [3] and are interesting for the large number (18) of normal subjects used.

REFERENCES

[1] F. Babiloni, F. Cincotti, L. Bianchi, G. Pirri, J. Millán, J. Mouriño, S. Salinari, and M.G. Marciani, “Recognition of Imagined Hand Movements with Low Resolution Surface Laplacian and Linear Classifiers”, Medical Engineering & Physics, 23(5), pp 323-328, 2001

[2] F. Babiloni, F Cincotti, L. Lazzarini, J. Millán, J. Mouriño, M. Varsta, J. Heikkonen, L. Bianchi and M.G Marciani,^.”Linear Classification of Low-Resolution EEG Patterns produced by Imagined Hand Movements”, IEEE Trans Rehabil Eng., 8(2), pp. 186-8, 2000.

[3] G. Pfurtscheller and C. Neuper, “Motor imagery and direct brain-computer communication”, Proceedings of the IEEE, 89(7): 1123–34, 2001