ECoG Real Time Signal Processing for Clinical Self paced BCI

Application

Nana Arizumi, Guillaume Charvet, Andrey Eliseyev, Jérémy Pradal, Serpil Cokgungor,

Nicolas Tarrin, Corinne Mestais, Tetiana Aksenova and Alim-Louis Benabid

CEA, LETI, CLINATEC

, MINATEC Campus, 17 Rue Martyrs, 38054, Grenoble Cédex, France

1 OBJECTIVES

The overall goal of the Brain Computer Interface

(BCI) project led at CEA/LETI/CLINATEC

is to

improve the quality of life of quadriplegic subjects.

BCI will allow them to control effectors such as an

exoskeleton, through recording and processing of the

electrical activity of their brain. To do this, a

wireless 64-channel ElectroCorticoGram (ECoG)

recording device WIMAGINE

(Wireless

Implantable Multi-channel Acquisition system for

Generic Interface with NEurons) has been designed

for long-term human implantation to interface an

electrode array to an external computer (Charvet et

al 2013). To decode the ECoG data, high resolution

algorithm has been constructed at CLINATEC

(Eliseyev et al., 2011); (Eliseyev and Aksenova

2013). Once the data are treated, they are used to

control the external effectors.

To reach the overall goal, it is crucial to

construct a whole software system working in real

time. In order to prepare the BCI software system

for the clinical trials, we demonstrated online real

time Electrocorticogram (ECoG) signal processing

using Monkey ECoG recordings corresponding to an

arm movement (Shimoda et al., 2012). The

algorithm of N-way Partial Least Square (NPLS)

regression family (Eliseyev and Aksenova, 2013) is

applied to extract linear model from the recordings.

The model is used to control the robotic arm JACO

(KINOVA) as a demonstrator.

2 METHODS

Figure 1 shows the schematic data flow for our BCI

system. The raw data should contain high temporal

resolution which does not limit to specific data

acquisition, e.g. EEG, MEG and ECoG.

Accumulation of data from several channels is

represented in Acquisition box in Figure 1. After the

data are collected in some buffer size, they are

mapped by the continuous wavelet transform (CTW)

to the temporal-frequency-spatial space (Acar et al

2008). Then they are sent to the linear prediction

model. To create the model, the algorithm of PLS

family (Eliseyev and Aksenova, 2013) is applied.

PLS is a statistical method for data analyses

particularly suited for high dimensional variables

(Geladi and Kowalski, 1986). PLS algorithms

provide stable linear models, which can then be used

to decode neuronal signal into commands for

external devices. Both CWT and prediction are

represented in Algorithm box in Figure 1. The data

transmission is represented in Effector application

box in Figure 1.

Figure 1: Flow of data decoding.

2.1 Specific Application

To test the system, publically available raw data are

considered (Shimoda et al., 2012), which contains

nine dimensional arm trajectory (shoulder, elbow

and wrist of x- y- z- coordinated captured by

VICON system) of Japanese macaque as well as

epidural ECoG signals of monkey’s brain (64

electrodes, sampling rate 1 kHz).

For decoding, block-wise Recursive N-way PLS

regressing is used (Eliseyev and Aksenova, 2013),

which can show correlation of 0.62, 0.80 and 0.85 of

shoulder, 0.54, 0.84 and 0.83 of elbow and 0.63,

0.85 and 0.82 of wrist for x- y- z- coordinates

Arizumi N., Charvet G., Eliseyev A., Pradal J., Cokgungor S., Tarrin N., Mestais C., Aksenova T. and Benabid A..

ECoG Real Time Signal Processing for Clinical Self paced BCI Application.

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

respectively. To form training tensor, 64 electrodes

ECoG signal and 1000ms window of analysis are

considered. The CWT with 84 frequencies between

0.6 and 300 Hz are performed with additional 100ms

tails using FFTW software (Frigo and Johnson,

2005). Then the signal was decimated in 100ms with

200ms sliding window. Using this training tensor,

predictive model is constructed. After the training

phase, the same features are considered for online

prediction. All of the computations are integrated

with OpenViBE (http://openvibe.inria.fr/) and

finally connected with the JACO robotic arm as

shown in Figure 2.

Figure 2: Comparison of predicted and actual movement.

3 RESULTS

To achieve real time, whole computation has to be

completed within the buffer size. With the specific

application from previous section, the temporal-

frequency-spatial dimension is 537600 and

predictive space has 9 degrees of freedom, namely

shoulder, elbow and wrist of x- y- z- coordinates.

With buffer size 100ms, the algorithm itself takes

83.81ms in average and simulating real time using

OpenViBE reach real time for more than 10 minutes.

4 DISCUSSION

The system can be applied to different algorithms

and data sets. The different model from PLS method

(Chao et al., 2010) is also tested. From the specific

applications, it is feasible to conclude that the model

using less than 64 channels, 84 frequencies and

1000ms window has decision rate at least 10Hz.

This is directly related to the CLINATEC BCI

project with ECoG signals of 64 channels using

linear predictive models.

ACKNOWLEDGEMENTS

The authors wish to thank the technical staff of

CLINATEC for their profound involvement in the

success of the project. The project received financial

support through grants from the French National

Research Agency (ANR-Carnot Institute), Fondation

Motrice, Fondation Nanosciences, Fondation de

l’Avenir, and Fondation Philanthropique Edmond J.

Safra.

REFERENCES

Acar, E., Bingol, C., A., Bingol., H., Bro, R., Yener, B.,

2007. Multiway analysis of epilepsy tensors.

Bioinform. 23 i10-8.

Chao, Z., C., Nagasaki, Y., Fujii, N., 2010. Long-term

asynchronous decoding of arm motion using

eletrocorticographic signals in monkeys. Front.

Neuroeng., 33.

Charvet, G., Sauter-Starace, F., Foerster, M., Ratel, D.,

Chabrol, C., Porcherot, J., Robinet, S., Reverdy, J.,

D’Errico, R., Mestais, C., Benabid, A-L., 2013.

WIMAGINE®: 64-channel ECoG recording implant

for human applications. In EMBC’13, 35

annual

international conference of the IEEE Engineering in

Medicine and Biology Society, July 3-7, Osaka, Japan.

Eliseyev, A., Aksenova, T., 2013. Recursive N-Way

Partial Least Squares for Brain-Computer Interface.

PloS one, In press.

Eliseyev, A., Moro, C., Costecalde, T., Torres, N., Gharbi,

S., Mestais, C., Benabid, AL., Aksenova, T., 2011.

Iterative N-way partial least squares for a binary self-

paced brain-computer interface in freely moving

animals. J Neural Eng. 2011 Aug;8(4):046012.

Frigo, M., Johnson, G., S., 2005. The design and

implementation of FFTW3. Proc. IEEE 93 (2), 216–

231.

Geladi, P., Kowalski, B., R., 1986. Partial least-squares

regression: a tutorial. Anal. Chim. Acta 185 1-17.

Shimoda, K., Nagasaka, Y., Chao, Z., C., Fujii, N., 2012.

Decoding continuous three dimensional hand

trajectories from epidural electrocorticographic signals

in Japanese macaques. J. Neural Eng., 9:036015.