2013 IEEE International Conference on Emerging Trends in Computing, Communication and Nanotechnology (ICECCN 2013)

227

EEG Signal Classification using Principal

Component Analysis with Neural Network in
Brain Computer Interface Applications
Kottaimalai R Pallikonda Rajasekaran M Selvam V Kannapiran B
II MTech C&I, Dept of ICE Prof. & Head, Dept of ICE Asst. Prof., Dept of ICE Asst. Prof., Dept of ICE
Kalasalingam University Kalasalingam University Kalasalingam University Kalasalingam University
Srivilliputtur, India Srivilliputtur, India Srivilliputtur, India Srivilliputtur, India
[email protected] [email protected] [email protected] [email protected]

Abstract - Brain Computer Interface (BCI) is the method of I. INTRODUCTION

communicating the human brain with an external device.
People who are incapable to communicate conventionally due The human brain is in fact a difficult system and
to spinal cord injury are in need of Brain Computer exhibits prosperous spatiotemporal dynamics. EEG is the
Interface. Brain Computer Interface uses the brain signals to method of obtaining the electrical activity of the brain as
take actions, control, actuate and communicate with the
analog signals using the instrument called
world directly using brain integration with peripheral
Electroencephalogram available in hospitals laboratory.
devices and systems. Brain waves are in necessitating to
eradicate noises and to extract the valuable features. The brain signals are taken using the techniques like
Artificial Neural Network (ANN) is a functional pattern invasive, partially invasive or non-invasive. In most of the
classification technique which is trained all the way through laboratory they are using the non-invasive technique to
the error Back-Propagation algorithm. In this paper in order acquire the brain signals by placing the electrodes on the
to classify the mental tasks, the brain signals are trained scalp using the conductive gel which improves the signal
using neural network and also using Principal Component strength. The received EEG signal is in microvolt and is
Analysis with Artificial Neural Network. Principal then amplified with the help of amplifiers. After
Component Analysis (PCA) is a dominant tool for analyzing
amplification, the amplified signal is given to computer
data and finding patterns in it. In Principal Component
through analog to digital converter for further processing.
Analysis, data compression is possible and it projects higher
dimensional data to lower dimensional data. By using Now the obtained signal is stored in the computer for
Principal Component Analysis with Neural Network, the doctor’s advice. In hospitals the use of EEG signal is in
redundant data in the dataset is eliminated first and the diagnosis of epileptic disorder, coma, brain death and
obtained data is trained using Neural Network. EEG data for encephalopathies. EEG also used for the diagnosis of
five cognitive tasks from five subjects are taken from the stroke, tumors and other main brain disorders, but this use
Colorado University database. Pattern classification is has lowered with the arrival of anatomical imaging
applied for the data of all tasks of one subject using Neural techniques with high spatial resolution such as Computer
Network and also using Principal Component Analysis with
Tomography (CT) and Magnetic Resonance Imaging
Neural Network. Finally it is observed that the correctly
(MRI).
classified percentage of data is better in Principal Component
Analysis with Neural Network compared to Neural Network Worldwide there are people who have
alone. malfunction in their motor activities due to paralysis or
having spinal injury because of accidents. They are
Keywords: Artificial Neural Network (ANN), Brain incapable to perform their own activities. A normal human
Computer Interface (BCI), Electroencephalogram (EEG), can interact with the computers using the peripherals like
Principal Component Analysis (PCA)
mouse, keyboard and etc. But the human with failed motor
function needs an interface to communicate to others or to
systems. For the patients having severe motor function
failure has care assistants to help them for their regular

978-1-4673-5036-5/13/$31.00 © 2013 IEEE

228

activities. Wide ranges of electronic device have been user performs left- or right-hand motor imagery to direct a
developed particularly to reduce the concern workload and left or right movement and performs foot imagery or
the number of care assistants needed. focuses on a flashing button to adjust the speed of the
Functional Electrical Stimulation (FES) is one simulated or real wheelchair.
procedure for the restoration of lost movement functions
[1]. Patients with spinal cord injury are incapable to III. SOFT COMPUTING TECHNIQUES
perform movements because the brain signal proposed for
progress of their extremities is not correctly transmitted to Soft computing is the combination of methodologies
the peripheral nerves. If these control signals could be (Neuro-Computing, Fuzzy Computing, Evolutionary and
absolutely given to Functional Electrical Stimulation Genetic Computing and Probabilistic Computing) intended
system, then the electric signals from the brain may be to model and make possible solutions to real world
capable to stimulate the peripheral nerve and respect the tribulations, which are not modeled or too complex for
command. Transmission of these signals to devices is mathematical modeling. Its aspire is to utilize the tolerance
essential but it is a time overwhelming work for severely for approximation (model features are similar to the real
paralyzed patients. ones but not the same), uncertainty (not sure that the
Brain–Computer Interfaces (BCIs) is the best model features belief are the same as that of the entity),
feasible way of providing the communication between the imprecision (model features quantities are not same as real
human and the system by means of brain signals [2]. By ones but close to them) and partial truth in order to achieve
using this BCI the patients can put across their views or close resemblance with human like decision making. The
needs by means of their brain signals just by thinking guiding theory of soft computing is to make use of these
process. The signal classification module is composed of tolerance to achieve, robustness tractability and low
the obtained EEG signal features extraction and the solution cost. Human mind is the role model for soft
transformation of these signals into device instructions. computing. Some of the soft computing techniques are
The EEG classification tactic depends on the inducement Artificial Neural Network (ANN), Fuzzy Logic (FL),
and, thereby, the reaction to detect motor imagery, event- Adaptive Neuro-Fuzzy Inference System (ANFIS),
related potentials, slow cortical potentials, or steady-state Principal Component Analysis (PCA) and evolutionary
evoked potentials. The predicted EEG drives the computation.
classification to some precise feature extraction methods.
1. ARTIFICIAL NEURAL NETWORK (ANN)
II. RELATED WORKS A neural net is an artificial illustration of the human
brain that tries to imitate its learning process. ANN is an
Kenji Nakayama et al [3] introduced efficient interrelated group of artificial neurons that uses a
pre-processing techniques in order to attain high mathematical model or computational model for
probability of exact mental task classification. The information processing [6]. ANN is a network of simple
preprocessing technique includes segmentation along time processing elements which can demonstrate complex
axis, amplitude of FFT of brain waves, reduction of overall performance, determined by the connections
samples by averaging and nonlinear normalization. between the processing elements and element parameters.
Charles W. Anderson et al [4] applied PCA independently ANN is an adaptive system that changes its structure based
to little segments of data and the origin vectors themselves on external or internal information that flows through the
are used as features for classification. In addition, time network. ANN computing approach to information
embedding the EEG by augmenting each sample with processing primarily involves a learning process with an
previous samples prior to PCA results in a representation ANN architecture that adaptively responds to inputs
that captures EEG variations in space and time. The according to a learning rule. After the NN has learned, the
resulting features are classified into categories trained network can be used to execute certain tasks
corresponding to which mental task a subject is depending on the exact purpose. The talent to learn by
performing in a brain-computer interface (BCI) paradigm. example and simplify are the principal characteristics of
Jinyi Long et al [5] introduced a hybrid BCI that uses the ANN. Classification of signals is done by using this ANN
motor imagery-based mu rhythm and the P300 potential to to obtain the correct classification percentage. ANN is
control a brain-actuated simulated or real wheelchair. The learned using the backpropagation algorithm in which the
229

errors for the units of the hidden layer are determined by a) Baseline task
back propagating the errors of the units of the output layer. b) Letter task
It is a systematic method of training multi-layer ANNs. It c) Mathematical task
contains an input layer, at least one intermediate/hidden d) Visual counting task
layer and an output layer in its network. Some of the ANN e) Geometric figure rotation task
learning parameters are Threshold, Goal, Epoch,
Sigmoidal function, Training type and Number of Hidden VI. IMPLEMENTATION AND RESULTS
layers.
1. NEURAL NETWORK CLASSIFIER
2. PRINCIPAL COMPONENT ANALYSIS (PCA) The brain signals taken from the above said seven
The graphical representation for higher dimensional electrodes which are all placed over the scalp are denoted
data is complex. PCA is an effective tool for analyzing as input layer for the neural network. [2] Here the five
data and finding patterns in it. It is a form of unsupervised different mental tasks classification are the outputs and the
learning and data compression is possible [7]. By output for baseline task is [1,0,0,0,0], for letter task is
projecting the data from higher dimensional into a lower [0,1,0,0,0], for mathematical task is [0,0,1,0,0], for visual
dimensional space that precisely characterizes the state of counting task is [0,0,0,1,0] and for geometric figure
the process, dimensionality reduction methods can rotation task is [0,0,0,0,1]. After applying various values
significantly simplify and progress process monitoring for the learning parameters like hidden layers, goal,
procedures. Principal Component Analysis is a threshold and epochs we found that the better
dimensionality reduction technique and it produces a classification is obtained for the learning parameters like
lower dimensional representation in a method that number of hidden layers as 15, epochs as 1000, threshold
conserves the correlation structure between the process as 0.5, goal as 0.01, sigmoidal function as logsig and
variables, and is best in terms of capturing the variability training type as trainscg. From the given dataset the data is
in the data. PCA projects high dimensional data to a lower taken from first trial of all tasks for one subject. Each trial
dimension and it projects the data in the least square sense, we obtained 2500 data. Therefore for all first trials of all
it captures big inconsistency in the data and ignores small five tasks we obtained 12500 data. The data is shuffled
inconsistency. well for mixing of data from all five tasks. Using the
neural network with backpropagation algorithm the first
IV. EEG DATA ACQUISITION 10000 data is trained well and then the next 2500 data is
used for testing. After the training function the result gives
In this paper, the EEG data used is collected from the the mean square error for testing and training, time taken
Colorado state University website [8] and was formerly for computation and the correctly classified percentage.
obtained by Keirn and Aunon using the following process.
The subjects were seated in an sound controlled room with 2. PRINCIPAL COMPONENT ANALYSIS (PCA) WITH
faint lighting. A non-invasive type electro-cap was wore to NEURAL NETWORK (NN) CLASSIFIER
record signals from positions P3, P4, C3, C4, O1, and O2 The signals obtained from the electrodes are
electrodes defined by the international 10-20 system for given to Principal Component Analysis for dimensionality
electrode placement. Data was recorded at a sampling rate reduction to remove the redundant variables in the data
of 250 Hz. Eye blinks were detected by placing a separate and the classified using Neural Network classifier with
channel of data observed from two electrodes placed over backpropagation. In Principal Component Analysis the
and under the subject’s left eye. better classification of signals is obtained for the learning
parameters like epochs as 1000, number of hidden layers
V. MENTAL TASKS as 3, goal as 0.01, sigmoidal function as tansig, threshold
as 0.5 and training type as trainscg.
There are five mental imaginary tasks taken from The brain signals are trained using Neural
seven subjects. During the tasks the data was recorded for Network and the training is shown in Fig.1. During the
10 seconds which contains 2500 data (i.e. for 1 sec=250 classification of the mental tasks using Neural Network
data) and tasks are repeated for 10 trials [9]. The tasks are classifier, the data is misclassified at the output ie., the
as follows. percentage of correct classification is low and is shown in
230

Fig.2. Similarly during the classification of the mental classification has been increased by using the Principal
tasks using Principal Component Analysis with Neural Component Analysis with Neural Network. Further the
Network classifier, the data is perfectly classified at the same dataset can be classified by using other soft
output ie., the percentage of correct classification is good computing techniques like Fuzzy logic, ANFIS in future
because of the reduction of the redundant variables in the for better results.
dataset and is shown in Fig.3.

Table 1: Comparison results of Neural Network and Principal

Component Analysis with Neural Network classifier.

PCA WITH
S.NO. PARAMETERS NN
NN

1 Hidden layers 15 3

2 Epochs 1000 1000

3 Goal 0.01 0.01

4 Threshold 0.5 0.5

5 Sigmoidal function Logsig Tansig

6 Training type Trainscg Trainscg

7 MSE Training 0.1099 0.1467

8 MSE Testing 0.1651 0.1478

Processing Time in
9 179.2919 87.0798 Fig.1: Neural Network Training
Seconds

Percentage Correctly
10 28.84 100
Classified

The comparison of the results of Neural Network

classifier and Principal Component Analysis with Neural
Network classifier is tabulated in Table 1 to show the
variation of mean square error during training, mean
square error during testing, computation time and the
percentage of correctly classified data for both type of
classification.

VII. CONCLUSION

A Neural Network has been applied to the BCI

problem. In order to improve accuracy of mental task
classification, several kinds of pre-processing is done to
generate the input data of the neural network. Compared Fig.2: Neural Network Classification Output implies the classification of
the mental tasks which are given as input to the Neural Network classifier
with the Neural Network model, the probability of correct
231

[5] Jinyi Long, Yuanqing Li, Hongtao Wang, Tianyou Yu, Jiahui Pan,
Feng Li, “A Hybrid Brain Computer Interface to Control the
Direction and Speed of a Simulated or Real Wheelchair” Journal of
medical signals and sensors 2011 Jul ;1 (3):206-13 22606677

[6] Hubert Cecotti and Axel Graser,”Convolutional Neural Networks for

P300 Detection with Application to BCI”, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol.33, NO.3 March
2011

[7] Kavita Mahajan, M. R. Vargantwar, Sangita M. Rajput

“Classification of EEG using PCA, ICA and Neural Network”
International Journal of Engineering and Advanced Technology
(IJEAT) ISSN: 2249 – 8958, Volume-1, Issue-1, October 2011

[8] Data from Colorado State University website

http://www.cs.colostate.edu/eeg/

[9] C. Anderson and Z. Sijercic “Classification of EEG Signals from

Fig.3: Principal Component Analysis with Neural Network Classification
Four Subjects During Five Mental Tasks” (1996) In Solving
Output implies the classification of the mental tasks which are given as
Engineering Problems with Neural Networks: Proceedings of the
input to the Principal Component Analysis with Neural Network
Conference on Engineering Applications in Neural Networks
classifier
(EANN'96), ed. by Bulsari, A.B., Kallio, S., and Tsaptsinos, D.,
Systems Engineering Association, PL 34, FIN-20111 Turku 11,
VIII. ACKNOWLEDGMENT Finland, pp. 407--414.

[10] Mahnaz Arvaneh, Cuntai Guan, Kai Keng Ang, and Chai Quek,
We thank the Department of Instrumentation and
“Optimizing the Channel Selection and Classification Accuracy in
Control Engineering of Kalasalingam University, EEG-Based BCI”, IEEE TRANSACTIONS ON BIOMEDICAL
(Kalasalingam Academy of Research and Education), ENGINEERING, VOL. 58, NO. 6, JUNE 2011
Tamil Nadu, India for permitting to use the EEG setup and
computational facilities available in Biomedical [11] Yuanqing Li, Jinyi Long, Tianyou Yu, Zhuliang Yu, Chuanchu
Laboratory which was setup with the support of the Wang, Haihong Zhang and Cuntai Guan, “An EEG-Based BCI
Department of Science and Technology (DST), New Delhi System for 2-D Cursor Control by Combining Mu/Beta Rhythm
and P300 Potential” IEEE TRANSACTIONS ON BIOMEDICAL
under FIST Program.
ENGINEERING, VOL. 57, NO. 10, OCTOBER 2010

VIII. REFERENCES [12] Chin-Teng Lin, Che-Jui Chang, Bor-Shyh Lin, Shao-Hang Hung,
Chih-Feng Chao and I-Jan Wang,“A Real-Time wireless Brain-
[1] Xicheng LIU, Shin HIBINO, Taizo HANAI, Toshiaki IMANISHI, Computer Interface system for drowsiness detection”, IEEE
Tatsuaki SHIRATAKI, Tetsuo OGAWA, Hiroyuki HONDA and Transactions on Biomedical Circuits and Systems, Vol.4, NO.4,
Takeshi KOBAYASHI “Construction of an Electroencephalogram August 2010
based Brain Computer Interface using an Artificial Neural Network”
IEICE TRANS. INF. & SYST., VOL.E86-D, NO.9 SEPTEMBER [13] Henrik Gollee, Ivan Volosyak,, Angus J.McLachlan, Kenneth J.
2003 Hunt, and Axel Graser, “An SSVEP-Based Brain–Computer
Interface for the Control of Functional Electrical Stimulation”,
[2] Gerwin Schalk and Eric C. Leuthardt, “Brain-Computer Interfaces IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING,
Using Electrocorticographic Signals”, IEEE REVIEWS IN VOL. 57, NO. 8, AUGUST 2010
BIOMEDICAL ENGINEERING, VOL. 4, 2011

[3] Kenji Nakayama & Kiyoto Inagaki, “A Brain Computer Interface

Based on Neural Network with Efficient Pre-Processing”,
December 2006 pg.673-676 International Symposium on Intelligent
Signal Processing and Communication Systems, Yonago
Convention Center, Tottori, Japan

[4] Charles W. Anderson and Jeshua A. Bratman, “Translating Thoughts

into Actions by Finding Patterns in Brainwaves” Proceedings of the
14th Yale Workshop on Adaptive and Learning Systems; New
Haven, CT, USA. 2008. pp. 1–6.