01615373
TRANSCRIPT
-
8/8/2019 01615373
1/4
AbstractBrain computer interface (BCI) algorithms areused to predict the torque generation in the direction of
shoulder abduction or elbow flexion using scalp EEG signals
from 163 electrodes. Based on features extracted from both
frequency and time domains, three classifiers are employed
including support vector classifier, classification trees and K
nearest neighbor. Support vector classifier achieves the highest
recognition rate of 92.9% on two able-bodied subjects in
average. The recognition rates we obtained on the able-bodied
subjects are among the highest compared with previous reports
on predicting motor intent using scalp EEG. This demonstrates
the feasibility of separating the shoulder/elbow torques usingscalp EEG as well as the potential of support vector classifier
in applications of BCI. Preliminary experiments on two
hemiparetic stroke subjects using support vector classifier
reports an accuracy of 84.1% in average, which shows an
increased difficulty in predicting intent presumably due to
cortical reorganization resulting from the stroke.
Keywords BCI, EEG, Support Vector Classifier,Shoulder Abduction, Elbow Flexion,
I. INTRODUCTION
Brain computer interface (or BCI) recently emerged as a
technique to convey brain derived signals for
communication and control [1-4]. These studies demonstrate
that BCI provides a non-muscular way to convey a usersintent. In particular, BCI based on electro-encephalogram
(EEG) is a promising candidate for use in human
rehabilitation. Compared with invasive methods that useneuronal action potentials or local field potentials recorded
within the brain, noninvasive EEG signals recorded from
scalp are convenient, safe and inexpensive. Variouscomponents of EEG signals have been found useful for
movement prediction including localized changes in spectral
power of spontaneous EEG related to sensorimotor processes [5][6], slow cortical potentials [7] and various
types of event-related potentials [8-10].However, despite the progress in using BCI for the
prediction of motor intents, the average recognition rate
achieved is around 80% [10-12]. Mensh et al. [13], as the
winner of BCI competition 2003, reported an 88.7%recognition rate for cursor movement in one able-bodied
subject. Whats more, there has been few exploration of
effective BCI algorithms that can separate motor tasksoriginating from close regions on the same hemisphere such
as those correspond to shoulder versus elbow movements.
Further study is thus required before BCI algorithms can be
applied for clinical use especially when brain reorganizationmay occur such as following stroke and spinal cord injury,
which may increase the difficulty of movement prediction.
EEG-based Discrimination of Elbow/Shoulder Torques using Brain Computer
Interface Algorithms: Implications for Rehabilitation
J. Zhou1, J. Yao
2, J. Deng
3, J. Dewald
2,3,4
1Department of Computer Science, Northern Illinois University, DeKalb, IL, USA2Department of Physical Therapy and Human Movement Sciences, Northwestern University, IL, USA
3Department of Biomedical Engineering, Northwestern University, IL, USA4Department of Physical Medicine and Rehabilitation, Northwestern University, IL, USA
In this paper, we present results of a preliminary study
on predicting the subjects intent of shoulder abduction
(SABD) or elbow flexion (EF) using advanced BCI
algorithms based on 163-channel scalp EEG signals. This isthe first attempt using BCI to separate shoulder/elbow motor
tasks in able-bodied and hemiparetic stroke subjects. It isalso the first time such a large number of electrodes are used
in EEG-based BCI study. In order to derive effectivealgorithms, we investigate the performance of combining
time-frequency feature extraction with three differentclassifiers, i.e., Support Vector Classier (SVC),
Classification and Regression Tree (CART), and K-nearestneighbor (KNN). Our preliminary results demonstrate that
SVC has the best potential in predicting upper arm
movements and is a promising candidate for the use of BCI
in rehabilitation applications.
II. METHODOLOGY
A. Experimental Setup
Each subject learned to self-initiate the generation ofisometric shoulder abduction (SABD) or elbow flexion (EF)
at a level of 25% of his/her maximum voluntary torques(MVTs). EEG and torques were collected during the
generation of isometric elbow/shoulder torque.
Subjects were cast at the wrist and secured to a six
degree of freedom (DOF) load cell with the shoulder at 75abduction, 40 flexion and the elbow at a 90 flexion angle.
In this position, the tip of the hand was approximatelyaligned with the median sagittal plane of the subject. In
order to minimize the effect of trunk muscle activation,
subjects were seated in a Biodex chair with the trunk
secured and the shoulders strapped to the back of the chair.A computer monitor was placed in front of the subject to
provide visual feedback of the torque generation in the
training protocol.Scalp recordings were made using a 163-channel EEG
system with active electrodes. The electrodes are mounted
on a stretchable fabric based on a 10/20 system. The capwas fitted on the head of the subject lining the Cz electrode
with the intersection of the planes defined by the nasion,
inion, and pre-auricular points. The skin under eachelectrode site was prepared and conductive gel was injected
Proceedings of the 2005 IEEEEngineering in Medicine and Biology 27th Annual ConferenceShanghai, China, September 1-4, 2005
0-7803-8740-6/05/$20.00 2005 IEEE. 4134
-
8/8/2019 01615373
2/4
to achieve electrode impedances lower than 5 k
throughout the experiment. EEG data were collected at 1000
Hz sampling rate. Anti-aliasing filtering (100 Hz) wasprovided before data acquisition. The system was equipped
active electrodes that provide a first amplification stage,
allowing detection of EEG signals with a higher SNR and
quicker preparation.
B. Feature Extraction
Both frequency and time domain features were
extracted from the EEG signal. First, a finite difference
Surface Laplacian (SL) transformation was applied to eachchannel as a spatial high-pass filter to reduce the smearing
effects caused by head volume conductor and to increase thesignal noise ratio (SNR). Due to the possibility that for the
same motor task subjects use different strategies, we used
principal component analysis (PCA) to calculate the
covariance complexity of the signals. When a data set was
considered highly complex, we divided the set into twosubsets using the average complexity value as the dividingline [14]. Next, we extracted features by decomposing
signals in both frequency bands and time courses. The
constant Q decomposition was conducted in the frequency
domain to divide the signals into frequency bins from 5 to34 Hz with a group of band-pass filters. Q is the ratio of
center frequency over bandwidth (set as 4). The informationin the time domain within each frequency bin is then
obtained by applying Hilbert transformation to extract the
profiles of the oscillatory activities. The profiles were
divided into equal-length (around 55 ms of each interval)with 50% overlapping. The feature for each time interval
was the integrated profile of that time duration, which can
be viewed as the instantaneous power on the correspondingtime-frequency grid. As the result of feature extraction, each
trial can be represented by a spatial pattern over the channel
distribution on the scalp where each channel has its two-dimensional feature (i.e., features in time and frequency
domains).
For more details of the exploited feature extractionalgorithm, refer to [10] and [14]. In this paper, our focus is
on comparing different classifiers, and particularly
examining the performance of support vector classifier inseparating the shoulder versus elbow movements in able-
bodied and hemiparetic stroke subjects.
C. Classifiers
a. Support Vector Classifier
The support vector classifier is a relatively new
classification technique developed by Vapnik [15] which
has reported strong performance in a number of real-worldapplications including BCI [12][16]. The basic idea is to
find an optimal separating hyperplane for two classes thatmaximizes the margin (see definition of margin below). The
algorithm looks for the optimization weight vector (w) and
offset (b) of the discriminating hyperplane by solving the
following quadratic problem:
minw,b
{|| w ||2 C i2
i
I
}
subject to,1)( iii bxwy
i
iiii Iixyw .,..,1,;0
where i are called slack variables which ensures that the problem has a solution when the data are not linearly
separable; yi is the ith label; i is the index of the sample
(vector); and i is the Lagrange coefficient obtained whenusing Lagrange theory to solve the above optimization
problem. The margin is defined as 2/||w||2. Geometrically, as
illustrated by Figure 1, the margin is the distance between
the bounding planes of the two data sets. The data vectorson the bounding planes are critical for finding the solution.
They are called support vectors, which explains the name ofthe algorithm.
To solve nonlinear problems, support vector classifier
conducts a transformation from input data space to featurespace by applying a kernel function to the data. Most
commonly used kernel functions are polynomial and
Gaussian. For details, see [17].
Fig. 1 Hyperplane and margin in support vector classifier.
b. Classification and Regression Tree
Classification and regression tree (CART) [18] is a
multi-stage binary decision tree. Every leaf node isassociated with a category. Every other node contains a
decision rule which partitions the feature space into two.The training process of CART typically contains the
stage of constructing a tree and pruning the tree to reduce
model complexity and avoid overfitting. The testing processwith CART is the process of following the decision rules of
the branches until a leaf node is reached and the category
associated with the leaf node determines the class label forthe testing sample.
4135
-
8/8/2019 01615373
3/4
As a classifier, CART has the advantage of doingimplicit feature extraction during the process of tree
construction. In other words, only the features are relevantwith the decision making are contained in the tree. However,
the decision boundaries of CART are linear and typically
parallel to the coordinates, therefore, its performance under
some conditions is less optimal.
c. K Nearest Neighbor
K nearest neighbor (KNN) is a simple classifier that
assigns a label same as the majority label among K nearest
neighbors of the testing sample. The distance measure between samples is application-specific and we used the
correlation coefficient between two feature maps of twotrials.
III. R ESULTS
A. Results from Able-bodied Subjects
Experiments were conducted on two able-bodied male
subjects both with right hand dominance. Each subject
performed around 100 trials of torque generation in eitherSABD or EF directions.
The evaluation of the algorithms consisted of two parts:training and testing. The training procedure provided each
time-frequency interval a weight w(t,f) [0, 1] proportionalto its contribution. It was calculated using the trainingrecognition rate achieved by each classifier on the particular
frequency/time grid. Following training, a testing procedure
was performed using 17 fold cross-validation. We dividedthe data into 17 sets with about equal numbers of trials per
set for each movement type. In each fold, one set was used
for testing and all the other trials were used to train theclassifier. The reported result is an average on 17 folds.
LibSVM 2.6 [19] is used in our experiment with default
parameters. For each fold, we calculated the recognition rateas Nc/N, where N is the number of total test trials and Nc is
the number of correctly predicted test trials.
The recognition rates for the two subjects are listed in
Table 1. Figure 2 depicts the weight distribution in thefrequency and time domains for the two subjects.
TABLE 1RECOGNITION RATE OF THREE BCI ALGORITHMS ON TWO SUBJECTS TO
DISCRIMINATE SHOULDER ABDUCTION AND ELBOW FLEXION.
Subjects\methods Nearest
Neighbor
Classifier
Classification
and
Regression
Tree
Support
Vector
Machine
N1 89.1% 87.6% 94.7%
N2 90.7% 82.1% 91.0%
Mean 89.9% 84.9% 92.9%
Figure 2 Weights on the frequency and time domain.
Our analysis resulting in the following observations::
1 The support vector classifier outperforms the other two
methods of CART and K Nearest Neighbor. Extracting
features from time and frequency domains makes it an
effective BCI algorithm allowing for an effective
separation between the intent to generate shoulder
abduction versus elbow flexion torques using scalp
EEG signals.
2 The reported recognition rate of 92.9% is among the
highest BCI performances among reports of EEG-based
motor classifications.
3 Figure 2 suggests that some frequency bands have
higher weights than others. However, the physiological
meaning of these differences in weights still requires
further exploration.
4 This paper reports one of the few experiments that uses
as many as 163 electrodes in EEG-based BCI. The
result confirms the feasibility and effectiveness of using
a large number of electrodes for signal collection. In
future work, we will seek to determine the optimalnumber of electrodes required to obtain the best
recognition rates.
B. Results from Hemiparetic Stroke Subjects
Obligatory coupling between certain shoulder andelbow torques is typically found in the moderately to
severely impaired arm stroke patients and results indiscoordination during movement. If movement intentions
can be successfully separated using BCI, then in the long
run, we may apply the algorithms to drive rehabilitation
interventions for the treatment of upper limb
discoordination. However, the associated challenge is theexpected brain reorganization following stroke [20], whichcan make the separation between shoulder and elbow
movement intentions more difficult.
To investigate the effectiveness of our algorithms and
the impact of an increased overlap in sensorimotor corticalareas used during elbow versus shoulder tasks following
stroke, experiments were conducted on two hemipareticstroke subjects.. Features were extracted from EEG signals
as described in Section II(B). Torques were separated using
the support vector classifier. Information regarding the two
4136
-
8/8/2019 01615373
4/4
subjects and preliminary results are listed in Table 2. Theerror rate is calculated as Ne/N, the recognition rate is
defined as Nc/N and the accuracy is defined as Nc/(N-Nr),whereNis the number of total test trials,Nc is the number of
correctly predicted test trials,Ne is the number of incorrectly
predicted test trials, andNris the number of rejected trials.
TABLE 2INFORMATION OF STROKE SUBJECTS AND DISCRIMINATION RESULTS
FOR SHOULDER ABDUCTION AND ELBOW FLEXION.SMA = Supplementary Motor Area; PM = Primary Motor. The results on eachsubject are averages of 17-fold cross validation.
Results\Subjects S1 S2 Mean
Age 60 51 -
Sex Male Female -
Dominant Hand Left Right -
Affected Hand Left Right -
Lesion Position R.
posterior
limb of IC
L. dorsal
lateral SMA
& PM
subcortical
white matter
-
Error Rate 12.5% 10.3% 11.4%
Recognition
Rate
48.0% 54.2% 51.1%
Accuracy 82.8% 85.3% 84.1%
As shown by Table 2, on one hand, the results show that
the reduced spatial resolution at sensorimotor cortices
increased the difficulty of prediction since the average errorrate (11.4%) is higher than that of able-bodied subjects
(7.1%); on the other hand, with the added rejection scheme,we achieve an average of 84.1% prediction accuracy on
impaired subjects. This demonstrates the feasibility of our
BCI algorithm in stroke patients and represents promising
preliminary results considering the challenge caused bybrain reorganization after stroke.
IV. CONCLUSION
In summary, this work demonstrates that support vector
classifier, together with frequency/time domain featureextraction, are effective algorithms for separating shoulder
abduction and elbow flexion intents, and thus suggests a
promising direction for EEG-based discrimination for otherupper limb movements as well as future applications in the
field of neuro-rehabilitation.
REFERENCES
[1] Wolpaw JR, McFarland DJ, Vaughan TM, and Schalk G. TheWadsworth Center brain-computer interface (BCI) researchand development program. IEEE Trans. Neural. Syst. Eng.11:204-207. 2003
[2] Pfurtscheller G, Neuper C. Mlle GR, Obermaier B, Krausz
G, Supp A. and Schrank C. Graz-BCI: State of the art and
clinical applications, IEEE Trans Neural Syst. Eng. 11:177-180, 2003
[3] Blankertz B, et al.. The BCI Competition 2003: Progress and perspective in detection and discrimination of EEG singletrials. IEEE Trans on Biomedical Eng 51:1044-1051, 2004
[4] Farwell LA and Donchin E. Talking off the top of your head:toward a mental prosthesis utilizing event-related brain
potentials.Electroencephalogr Clin Neurophysiol. 70(6):510-
23. 1988.[5] Wolpaw JR, McFarland DJ, Neat GW, and Forneris CA. An
EEG-based brain-computer interface for cursor control. Electroencephalogr Clin Neurophysiol78: 252-259, 1991
[6] Pfurtscheller G and Neuper C. Motor imagery activatesprimary sensorimotor area in humans.Neurosci Lett239: 65-68, 1997
[7] Birbaumer N, Kbler A, Ghanayim N, Hinterberger T,
Perelmouter J, Kaiser J, Iversen I, Kotchoubey B, Neumann N, Flor H. The thought translation device (TTD) forcompletely paralyzed patients. IEEE Trans Rehabil Eng
8:1903. 2000.[8] Donchin E, Spencer KM, Wijesinghe R. The mental
prosthesis: assessing the speed of a P300-based brain-computer interface. IEEE Trans Rehabil Eng; 8(2):1749.2000.
[9] Middendorf M, McMillan G, Calhoun G, Jones KS. Brain-
computer interfaces based on the steady-state visual-evokedresponse.IEEE Trans Rehabil Eng; 8(2):2114. 2000.
[10] Wang T, Deng J, and He B. Classifying EEG-based motorimagery tasks by means of timefrequency synthesized spatial
patterns. Clinical Neurophysiology 115:27442753, 2004.
[11] Lal TN, Schrder M, Hingerberger T, Weston J, Bogdan M,Birbaumer N, and Schlkopf B. Support Vector ChannelSelection in BCI.IEEE Trans on Biomedical Eng51(6):1003-1010, 2004
[12] Garrett D, Peterson DA, Anderson CW, and Thaut MH.Comparison of linear, nonlinear, and features selection
methods for EEG signal classification. IEEE Trans on NeuralSystems and Rehabilitation Engineering11(2):141-144, 2003
[13] Mensh BD, Werfel J and Seung, HS. BCI competition 2003-data set Ia: combining gamma-band power with slow cortical
potentials to improve single-trial classification of
electroencephalographic signals. IEEE Trans on Biomedical
Eng51:1052- 1056. 2004.[14] Deng J, Dewald JP and Yao J. Classifying Isometric
Shoulder/Elbow Motor Tasks Using Time-frequency
Synthesized Spatial Patterns of Scalp EEG signals. Inpreparation.
[15] Vapnik V. The nature of statistical learning theory. Springer-Verlag. New York. 1996.
[16] Zhou J and Pasteris G. Data driven approach to designing
minimum Hamming distance polychotomizer. Proceedings of
ACM Symposium of Applied Computing, pp27-31, 2005[17] Burgers CJC. A tutorial on support vector machines for
pattern recognition. Knowledge Discovery and Data Mining2(2):121-167, 1998
[18] Breiman L, Friedman J, Olshen R, and Stone C Classification
and Regression Trees. Chapman & Hall, New York. 1984[19] http://www.csie.ntu.edu.tw/~cjlin/libsvm/[20] Yao J and Dewald J. Muscle Focus: A New Biomechanical
Model-Based Index And Its Application In Quantifying TheMuscle Coactivation Patterns During Isometric TorqueGeneration At The Elbow And Shoulder. EMBC, SanFrancisco, 2004
4137