8/13/2019 05109322

1/4

DEVELOPMENT OF BRAIN-COMPUTER

INTERFACE (BCI) MODEL FOR REAL- TIME

APPLICATIONS USING DSP PROCESSORS

G.Karthikeyan

Undergraduate student,Department of Biomedical Engineering

SSN College of Engineering, Chennai 603110Email: g.karthikeya@gmail.com

N.Sriraam

Center for Biomedical Informatics and SignalProcessing, Department of Biomedical Engineering

SSN College of Engineering, Chennai 603110Email: sriraamn@ssn.edu.in

Abstract: Brain computer interfaces (BCIs) are systemsthat provide translation of the electrical activity of the brain into

commands which can control devices in real-time. It is known

that most BCI based application involves non-invasive EEG

signals in combination with machine learning algorithms for

control of devices. Although there has been some real- time

implementation of BCI including cursor movements controlled

by a persons thoughts, they have not been done as a portable

device, but rather as a laboratory experiment controlled using

multi- core processors and computers. This paper focuses on

real- time implementation of BCI for motor imagery tasks using

a floating point DSP processor (TMS320C6713DSK). The

proposed work involves FIR based preprocessing filter to

remove extraneous noise including Electrooculogram (EOG) and

Electromyographic (EMG) signals followed by feature extraction

using time-frequency domain features. The implications of this

kind of a development in Brain- Computer Interface would be

tremendous since the whole system can be made into a portable

device or a System on a Chip (SoC).

Keywords Brain- Computer Interface, DSP Processor,Neural Networks, Linear Discriminate Analysis (LDA).

I. INTRODUCTIONBrain-computer interface (BCI), a rapidly developing

technology, has been constantly improving and new

techniques being constantly added to the already existing ones.The main goal of Brain- Computer Interface is creating a new

pathway of communication for persons with severe motordisabilities [1-2,4-5]. Most of the present day BCI researchis carried out using Electroencephalogram (EEG) signals

owing to its non- invasive nature. [1-2]. Though EEG ispredominantly diagnostic in nature, their inherent uniqueness

in different features can aid in development of BCI devices.This paper focuses on the real- time application of BrainComputer Interface by using a DSP processor for filtering the

raw EEG data and using this filtered data for feature extractionand classification.

II. MATERIALS AND METHODSA. EEG Data

Figure 1 Time sequence specification considered for datarecording

The data set used in this study was recorded at theBiomedical Instrumentation Laboratory of the SSN College of

Engineering; Chennai, India. The recording procedure isdescribed below:

The electrodes were placed with accordance to the 10-20electrode placement system and the recordings were donefrom a single subject for three sessions with each session

yielding 300 usable trials and recordings from each of theconsidered channels were sampled at a rate of 256 Hz. Fig. 1

details the recording time sequence followed.The characteristic of the filter used was that of a band pass

filter with the filter range being 0.1Hz to 100Hz. The

impedance value of the recording set up was limited within 20K. The channels recording the frontal activities were

considered for the study as the frontal electrodes are the mostsensitive ones for the recording of imagery motor activities

Different combinations of frontal electrodes were taken intoaccount and the study conducted and the results evaluatedaccordingly. The frequency range of the signals used in the

study was that of the characteristic -waves of the imaginarythought processes. Fig. 2 shows the sample recordings.

FrD5.4Proceedings of the 4th InternationalIEEE EMBS Conference on Neural EngineeringAntalya, Turkey, April 29 - May 2, 2009

978-1-4244-2073-5/09/$25.00 2009 IEEE 419

8/13/2019 05109322

2/4

III. EEG FILTER DESIGN COMPARISONFig 3 shows the overview of the proposed process. Filter

design is done using both MATLAB and the DSP processorand the EEG data was filtered using the filter designed and the

filter characteristics of both the filters were found to be the

same. The filter transfer function has been given in Fig 4. The

Cut- Off frequencies for the Band- Pass filter of order 400 wasgiven as 1 Hz and 30 Hz. Fig 4 shows the magnituderesponse of the filter designed.

0 10 20 30 40 50 60 70 80 90-2000

-1000

0

1000

2000

Time (seconds)

Amp(microvolts)

Trial 1 right hand signal

0 10 20 30 40 50 60 70 80 90-2000

-1000

0

1000

Time (seconds)

Amp(microvolts)

Trial 1 left hand signal

Figure 2 Plot of the data sets of both right and left handsused in the study

A. TMS320C6713DSK- Overview

The C6713 DSK is a low-cost standalone developmentplatform that enables users to evaluate and develop

applications for the TI C67xx DSP family. The DSK alsoserves as a hardware reference design for the TMS320C6713

DSP.

Figure 3 Block set up for the study

The TMS320C6XXX processors are based on the Very LongInstruction word (VLIW) architecture. This architecture

facilitates development of a high- level language compiler.The internal program memory is structured so that a total of

eight instructions can be fetched every cycle.

Features of the C6713 include 264 kB of internal memory(8kB as L1P and L1D Cache and 256kB as L2 memory shared

between program and data space), eight functional or

execution units composed of six arithmetic-logic units (ALUs)and two multiplier units, a 32-bit address bus to address 4 GB(gigabytes), and two sets of 32-bit general-purpose registers.Fig.5 shows the typical DSP flow process.

IV. FEATURE EXTRACTIONFeature Extraction was done using time- domain, frequency

domain and non- linear chaotic parameters. The featureextraction methods used were, Levinson- Durbin (LD)

method, power spectral density using the Burg (PSD Burg)method and the Hurst exponent (HE) method. The LDmethod is used to calculate the autoregressive (AR)

Coefficients of an nth- order AR linear process is calculatedby using the relation:

H(z)=1/A(z)=1(1+a(2)z^1+....+a(n+1)z^-1) ->(1)

Hurst exponent which possesses spectral characteristics is oneof the key parameters in non- linear dynamics. The values of

HE is obtained using the statistical technique called Rescaledrange analysis [3]. Figs 6-7 show the feature extraction resultsobtained for the test data considered for this work.

Figure 4 Magnitude Response of the FIR filter (DSK o/p)

V. CLASSIFICATIONClassification of the extracted features were carried out usingfeedback (Elman Network) and feedforward (multilayer

perceptron) neural network classifiers. The results obtained

using the classifiers were compared and the most suitableclassifier for a given set of feature extraction techniques was

studied. The performance of the classsifiers were evaluatedin terms of classification accuracy (CA) which is defined as

in (2)

420

8/13/2019 05109322

3/4

CA= CD/TA (2)

Where,

CD refers to total number of correctly detected lefthand and right hand patterns by the neural network

TA refers to total number of applied patterns.For this study, 400 features are used for training and 600 fortesting.

Figure 5 DSP process flow

Figs 8(a)-(c) shows the classification results obtained usingElman neural network. It can be observed an overall

classification accuracy of 99.5% is achieved using the Hurstexponent with Elman neural network.

0 5 10 15 20 25 30 35 40 45 500

0.005

0.01

0.015

0.02

0.025

0.03

EEG patterns

Levinson-DurbinCoefficients

Imaginary LH Movement-Levinson

Imaginary RH Movement-Levinson

Figure.6 Feature extraction results using Levinson- Durbin

method

VI. CONCLUSIONThis paper discusses the design and comparison of

preprocessing filters for BCI applications using real- time DSP

processors. The different types of features extracted wereclassified using Neural Network classifier. For the filterdesigned using the DSP processor, it was found that the

transfer characteristics of the filter was quite similar to thetransfer characteristics of the filter designed using MATLAB.

The work carried out on MATLAB can be extended to the

DSP processor and a fully functional real- time BCI systemcan be developed based upon the preliminary works proposed.

0 5 10 15 20 25 30 35 40 45 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

EEG patterns

PSD-Burg

Imaginary LH Movement-Pburg

Imaginary RH Movement-Pburg

Figure 7 Feature extraction results using PSD Welch Method

0 100 200 300 400 500 6000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

No of samples

EN

output

Elman network with Levinson

Figure 8(a): Classifier output using Levinson- Durbin

0 100 200 300 400 500 6000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

No of samples

EN

output

Elman network with Pburg

Figure 8(b): Classifier output using PSD Burgs Method

421

8/13/2019 05109322

4/4

0 200 400 600 800 1000 12000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

No of samples

EN

output

Elman network with HE

Figure 8(c): Classifier output using HE

REFERENCES

[1 ]N.F.Incea,A.H.Tewfika, S.Aricab,Extraction subject-specific motor

imagery timefrequency patterns for single trial EEG classification,Computers

in Biology and Medicine 37 (2007) 499 508

[2] J.Wolpaw. Brain-computer interfaces as new brain output pathways. JPhysiol, 579:613--619, 2006.

[3] Bassingthwaighte JB, Raymond GM, Evaluating rescaled range analysis

for time series, Ann Biomed Eng 1994 Jul-Aug; 22(4):432-44.

[4] An adaptive filter bank for motor imagery based Brain Computer

Interface, Thomas, Engineering in Medicine and Biology Society, 2008.

[5] J. Wolpaw, N. Birbaumer, D. McFarland, G. Pfurtscheller, and T.

Vaughan. Brain-computer interfaces for communication and control. (Invited

review). Clin Neurophysiol, 113:767--791, 2002.

[6] V.Srinivasan, C.Eswaran, and N.Sriraam. 2005. Artificial Neural Network

based epileptic detection using Time-domain and Frequency-domain features ,

Journal of Medical Systems, Kluwer Academic Publishers, New York, USA

Volume 29, Number 6. 647-660.

[7 ]N.J. Huan and .Palaniappan,Electroencephalogram Signal Classification

Using Linear Discriminant Analysis for Brain-Computer Interface Design,

Proceedings of M2USIC 2004, 9-12 ,2004.

[8] N. Sriraam, G. Karthikeyan, J. Kamala Kannan Performance evaluation of

imaginary motor activities for Brain- Computer Interface Applications.

International Symposium on Global Trends in Biomedical Informatics,

Chennai, Jan 2008.

[9] Model-based responses and features in Brain Computer Interfaces,

Kamrunnahar, Engineering in Medicine and Biology Society, 2008. EMBS

2008.

[10] Classification of EEG with structural feature dictionaries in a braincomputer interface,Goksu, Engineering in Medicine and Biology Society,

2008. EMBS 2008.

[11] An Intelligent Brain Computer Interface of Visual Evoked PotentialEEG, Chen, Intelligent Systems Design and Applications, 2008. ISDA '08.

[12] Linear Discriminant Analysis on Brain Computer Interface, Perez,

Intelligent Signal Processing, 2007. WISP 2007.

[13] Non-invasive classification of cortical activities for brain computer

interface: A variable selection approach Besserve, M, Biomedical Imaging:From Nano to Macro, 2008. ISBI 2008.

[14] A DSP based Brain Computer Interface system, Morrison, Masters

Thesis, University of Florida.

[15] Digital Signal Processing and Applications with the C6713, Ralph

Chassaing.

[16] Bill Davidson, Matlab program for Hurst exponent, October 2003

422