2009 5th International Colloquium on Signal Processing & Its Applications (CSPA)

Brain Machine Interface: Motor Imagery

Recognition with Different Signal Length
Representations
Hema C.R., Paulraj M.P., S.Yaacob, A.H. Adom and R. Nagarajan
School of Mechatronic Engineering
Universiti Malaysia Perlis, Jejawi, Perlis, Malaysia
hema@unimap.edu.my

Abstract- This work investigates how signal representations
affect the performance of a motor imagery recognition system,
specifically we investigate on recognition accuracy and
computational time of a brain machine interface designed using
motor imagery. Experiments show that the signal length should
not be larger than a critical range for good recognition accuracy.
The results presented here is a part of our work on the design
and development of a brain machine interface to operate a
wheelchair. EEG motor imagery signals recorded from the motor
cortex area using non-invasive electrodes, are used for
recognition of four tasks namely, left, right, forward and stop.
Experiments are conducted for 12 signal representations from
signal lengths varying from 3s to 0.25s. From the results it is
observed that good recognition accuracies (93.2% -94.2%) are
obtainable for 2s to 3s signal representations
I. INTRODUCTION
Brain Machine Interfaces (BMI) are devices which use the
power of thought of the human brain to drive external devices
such as robotic limbs and wheelchair. This concept once
possible only in the realm of science fiction has become a
reality with the knowledge on the working of the brain and
relation between neural activity and hand movements. The
ambitious goal is to provide tetraplegic and patients suffering
from a variety of severe motor disorders such as paralysis and
amyotrophic lateral sclerosis with the means to act and
communicate by replacing the control of muscles with the
control of artificial devices through brain activity. In this
paper we present an analysis to determine the optimal signal
length required for direct translation into a control signal to
operate a wheelchair.
Literature shows that Motor Imagery classification for
BMI systems, use signal lengths varying form 1.25s to 10s [1-
6], as most systems are based on synchronous BMI, where
predefined time cues are given to the subject to imagine motor
tasks. However in real time asynchronous BMI, choosing the
optimal signal length used for classification can improve the
computational time and hence the translation time of the
system. In this paper an analysis is conducted to determine the
signal length of motor imagery signal required in an
asynchronous BMI to achieve good task classification. Here
the motivation is to reduce the computation complexity and
time to improve the bit rate transfer of the control signal to the
wheelchair.
II. SIGNAL LENGTH IN FEATURE EXTRACTION
A. Protocol and Data Collection
In this analysis EEG motor imagery signals collected in
our previous work [7] is used. Data from one healthy subject
is used in this experiment. The subject performs four motor
imagery tasks namely; relax, forward, left and right. Each
recording lasts for 10 s. EEG is recorded using two gold plated
cup electrodes placed at the C3 and C4 locations on the motor
cortex area as per the International 10 -20 Electrode Placement
System [8], The EEG signals are amplified and sampled at 200
Hz using an ADI Power lab amplifier. The signals are
recorded for twenty trials per task. In our analysis the 10s
length signals are split into different signal lengths ranging
from 3s to 0.25s, thus 12 signal lengths were analyzed to
determine an optimal signal length.
B. Feature Extraction
In the first analysis 3s signals are considered for
classification. The 3s motor imagery data are filtered using a
chebyshev band pass filter and band width signals are
extracted for frequency bands 8-10Hz, 10-12Hz, 12-15Hz, 15-
18Hz, 18-30Hz and 30- 44Hz , the sum of the band power
values are extracted and a logarithmic transform is performed.
These features are used to model a recurrent neural network.
For each signal length a neural network model is created; the
analysis is repeated for other lengths, hence twelve such
models are analyzed.
C. Classification
The signal length data are classified using an Elman
recurrent neural network [ERNN]; these are dynamic networks
with feedback connections which add the ability to also learn
the temporal characteristics of the data set. In this study
ERNN architecture with three layers is used. The ERNN
makes a copy of the hidden layer which is referred to as the
context layer. The purpose of the context layer is to store the
pervious state of the hidden layer at the previous pattern
presentation [9]. This improves the classification rate and
training time of the network in comparison to a feed forward
neural network.
A multilayer ERNN with one single hidden layer is trained
by the BP algorithm to classify the four states represented by

978-1-4244-4152-5/09/$25.00 ©2009 IEEE 37

the EEG features. Each network model has 6 input nodes and International Conference of the IEEE EMBS, USA, Sep 2004, pp 4359
-4362.
4 output nodes, the hidden layer nodes are chosen
[5]. N-J Huan and R. Palaniappan, “Classification of Mental Tasks using
experimentally as 9. The learning rate is chosen to be 0.001. Fixed and Adaptive Autoregressive Models of EEG Signals,” Proc. of
240 data samples are used in this experiment for each model. 26th Annual Intl. Conf. IEEE EMBS, CA, USA, pp 507 – 510, 2004.
The training and testing error tolerances are 0.001 and.05 [6]. Z.A Keirn and J I. Aunon, “A New Mode of Communication between
respectively. The training and testing data are normalized Man and his Surroundings” IEEE Trans. Biomedical Engineering, Vol.
using a binary normalization. The ERNN is trained with 80% 37.no. 12. pp 1209-1214, 1990.
data set. Selections of the training data are chosen randomly. [7]. Hema C.R., Paulraj M.P., S.Yaacob, A.H.Adom, R.Nagarajan, “Motor
Imagery Signal classification for a Four State brain Machine Interface”,
Training is conducted until the average error falls below 0.001 International Journal of Biomedical Sciences , Vol 3 No. 1 Winter 2008,
or reaches a maximum iteration limit of 1000. pp 76 -81.
[8]. D.G. Domenick, “International 10-20 Electrode Placement System for
III. RESULTS Sleep”, 1998. http://members.aol.com/aduial/1020sys.html
[9]. Andries P. Engelbrecht,” Computational Intelligence an
Computational time and recognition accuracy is of prime Introduction”, John Wiley and Sons Ltd., 2002.
concern in real time BMI systems, experiments were
conducted to find the minimum signal length required from a
motor imagery to recognize the task performed. The TABLE I
TYPE SIZES FOR PAPERS
recognition results for various signal lengths is shown in Table
I, results show that good accuracy is achievable for signal
lengths varying from 2s to 3s, lower lengths did not perform Average Classification
EEG Signal Length
Accuracy
well. Offline experiments on single motor imagery trials were in seconds
%
conducted to validate the required signal length; the
0.25 80.2
experiment results had recognition accuracy of 93.2%.and
above with computational time varying from 2 to 9 seconds. 0.5 83.8

0.75 80.5
IV. CONCLUSION
1.0 71.1
This paper presented an analysis to determine the optimal 1.25 65.9
signal length for recognition of a motor imagery task signal
1.5 88.8
for real time BMI systems. Experiments are conducted on 12
data sets with varying signal lengths from 0.25s to 3s. Results 1.75 92.7
validate that signals length from 2s to 3s is sufficient to
2.0 93.2
obtained good recognition accuracy. The analysis results show
that better performance for lower signal lengths are possible as 2.25 94.1
against the traditional signal lengths of 4s and above used in
2.5 93.6
many research works on BMI. In our future works real time
experiments will be conducted using the results of this 2.75 93.7
analysis to implement a real time BMI for wheelchair control. 3.0 94.2

ACKNOWLEDGMENT

The authors wish to thank the MOSTI, Malaysia for the

FRGS Grant 9003 00187 which partly funded this research

REFERENCES

[1]. C. Guger, H. Ramoser, G Pfurtscheller, “ Real-Time EEG Analysis with

Subject Specific Spatial Patterns for a Brain-Computer Interface”, IEEE
Transactions on Rehabilitation Engineering Vol.8, No. 4, December
2000, pp 447 to 456.
[2]. E.B. Sadeghian, M.H. Moradi, “Continous Detection of Motor Imagery
in a Four Class Asynchronous BCI”, Proc. of 29th Annual International
conference of the IEEE EMBS, Lyon, Aug 2007, pp 3241 -3244.
[3]. Stefan Cosochi, Rodica Strungaru, Alexandra Ungureanu,Mihaela
Ungureanu, “ EEG Feature Extraction for Motor Imagery”, Proc. of 28th
Annual International conference of the IEEE EMBS, USA, Sep 2006,
pp 1142-1145.
[4]. Tao Wang, Jie Deng, Bin He, “ Classification of Motor Imagery EEG
Patterns and their Topographic Representations”, Proc. of 26th Annual

38