Non-Invasive EEG Feature extraction and
recognition of epileptiform activity using Adaptive
resonance theory
Aswathappa.P 1 , S.Ranjitha2 Dr .H. N. Suresh3
3
Abstract : The recent research in artificial neural networks
(ANN) have provided us many information about a natural
learning strategy. Artificial neural network consists of densely
interconnected simple neuron like elements. Information is
represented in the strengths of the connections (synapses)
between elements and is processed in parallel. Different neural
network models are specified by the topology of connections,
characteristics of processing elements and training rules. The aim
of this study is to develop a simple model for characterization of
Epileptical seizure. Emerging Adaptive Resonance Theory
(ART2 Neural Network) are used in this study to detect and
generate epileptical seizure.
Keywords: Adaptive Resonance Theory,
networks (ANN, epileptical seizure.
artificial
neural
ASWATHAPPA P
Asso. Professor
Dept. of IT, BIT Bangalore
I.INTRODUCTION:
In this study, adaptive resonance theory (ART 2) neural
network are investigated for on-line unsupervised recognition
(spikes for epilepsy monitoring) and also epileptical seizures
are generated from output nodes of the network. ART 2
networks are self organizing systems that clusters data into
different classes. Learning is completed when these classes are
labeled by an expert and are saved in a look-up table.
Recognition of new data is accomplished by finding the class
assigned by ART 2 and searching through the table to get the
proper labels. Unrecognized inputs are put into a new class for
future labeling. In this study an ART 2 neural network with 32
inputs was developed and trained using EEG data containing
spike and non-spike waveforms. For comparison a 32 input
multilayer perceptron is also constructed and evaluated for
comparison.
Aswathappa.P 1
1
Research scholar, Dept. Of ECE, Karpagam University,
Coimbatore, Tamil nadu, , India.
paswathappa@yahoo.com1
S.Ranjitha2
Final BE(ECE), Bangalore Institute of Technology,vv puram
,Bangalore-04,India
2
Dr .H. N. Suresh3
Professor, Bangalore Institute of Technology, Dept.of
Elecronics and Instrumentation Engg.,
Bangalore04,Karnataka,India
hn.suresh@rediffmail.com, 2
Evaluation with three sets of EEG patterns indicates that the
ART 2 Neural Network's performance is better than that of
multilayer perceptron showing its high potential. Considering
that ART 2 can be trained with one or few iteration (compared
to thousands required for back propogatation networks), hence
this can used for on line trainining. Such systems are required
for long-term EEG monitoring due to large variations of spike
waveforms that occur among individual patients during long
recording time periods.
Many researchers tasks include the recognition and
generation of abnormal electrical wave patterns. Traditional
algorithmic methods are insufficient to characterize epilepsy
seizure. Medical diagnosis is one of the prime examples of this
problem. Although many expert systems or so called
"intelligent" devices have been developed using traditional
Artificial Intelligence (AI) approaches. Intelligent systems are
developed using symbol and rule based AI techniques. In
essence, traditional AI requires the existence of a domain
expert who knows the answer to many questions and can
specify the rules to a knowledge engineer. This is however,
hardly the case in medicine since most of our knowledge is
incomplete. This research focuses on the problem of
epileptical EEG spike detection for the diagnosis and to
construct an ART memory to evaluate epilepsy.
Although many rules are proposed for the detection of
spikes (e.g. width, amplitude, sharpness, etc.) they are not up
to the satisfaction. Many EEG technicians are not aware of
these rules consciously, yet they can identify spikes in
ongoing EEG easily and accurately. Humans learn to
recognize EEG spikes by seeing a number of such waveforms
and generalizing their characteristics. Better success can be
expected if a natural learning strategy is used by humans and
adapted in machines. Multilayer perceptrons trained with back
propagation (BP) [1] are the best known and most utilized NN
models. Pattern recognition and classification by neural
networks have several major advantages over traditional
methods. Neural networks use massively parallel, nonalgorithmic processing which are in some ways similar to
those used by the human brain. As a result, many competitive
hypotheses have been explored simultaneously. Neural
networks store the patterns presented to them by modifying
their states. Pattern elements are not memorized individually
hence a large numbers of complex patterns can be stored.
Page 1 of 6
Neural networks have other advantages over current
traditional systems that they are fault-tolerant and relatively
insensitive to the underlying statistical distributions of patterns
to be recognized. Neural network applications in EEG have
been primarily focused on sleep staging, anesthesia
monitoring and transient event detections such as spikes and
K-complexes [2,3].
All these studies used multilayer
perceptrons to train various EEG waveforms. These studies
have shown that transient EEG events such as spikes and Kcomplexes can be successfully recognized [4]. These networks
are trained off line with a fixed set of labeled data using the
backpropagation algorithm and do not adapt to new conditions
during monitoring. Retraining with new patterns generally
degrades the performance of the network in recognizing
previously learned patterns. This property of the multilayer
perceptron results in a rigid, non adaptive system that cannot
be modified or customized.
To solve this problem new neural network architecture,
which adapts to new circumstances is needed. Once the
pattern recognition task has been completed then the network
behavior can be modified by new information. As new
patterns are added to the knowledge base (e.g., as data become
available from more patients) the network may adapt itself to
improve its recognition rate. Adaptive Resonance Theory
(ART) architecture was originally developed by Grossberg
and Carpenter [5]. The main thrust of this research is to
develop on line adaptive methods for the detection of EEG
spikes using ART neural networks. A brief description of
ART neural networks is presented in this section. We have
developed techniques for the detection of EEG spikes. These
techniques can be used for long term monitoring of seizure
patients by continuously training the system to recognize new
waveforms on line. After development of these techniques,
this can be used for the detection of other transient EEG
waveforms and biomedical signals.
input patterns. This type of Neural Network architecture has
the ability to learn new patterns in one or a few iterations
without forgetting the old ones (learned patterns). This
property of adapting to new information makes ART 2 NN
ideal for on line training with individual patients for long term
monitoring, since EEG waveforms obtained from each patient
are different. Originally two ART networks, ART l and ART
2, have been developed. ART 1 and ART 2 differ essentially
in the nature of input patterns. ART l network works only
with binary inputs while ART 2 works with analog patterns or
binary inputs. In this research we are using ART 2 networks
since EEG spike patterns are analog in nature. A detailed
diagram of the ART 2 model used in this study is shown in Fig
2.1 .A short description of the network is given below. A more
detailed description of the network can be found in [6]. As
seen in Fig. 2.1, ART2 neural network consists of two major
parts attentional subsystem and orienting subsystem.
Orientingsubsystem and Attentional sub system
II. ART 2 Neural Network
Architecture
ART neural networks are unsupervised, self organized,
and self stable types of NN architecture. ART is the type of
architecture that remains plastic (adaptive) in response to new
input patterns and yet remains stable, i.e., it preserves its
previous learned patterns. This Neural Network models
function as follows. First an input pattern is given to the
network on a feed forward (bottom-up) basis, which produces
a feed backward pattern (top-down) of activations, over one of
the existing node classifier. These feed backward and feed
forward patterns are then compared. If a match occurs
(resonance) within specified limits (vigilance), then this
pattern is classified as a member of that winning node. If not,
the remaining nodes are searched, failure to find an
appropriate node results in the formation of a new node for
this new pattern.
ART type neural networks are particularly suited for
pattern classification problems that cannot be defined by static
Fig. 2.1 - ART2 system diagram
The attentional subsystem consists of F1 and F2 layers. All the
nodes in F1 and F2 are fully interconnected with each other and
the weights associated with them are in the form of long-term
memory (LTM). F1 layer is composed of many sub layers
whose interconnections are depicted in Fig. 2.1.All the
processing elements in ART 2 models are governed by the
neuronal equation.
€ d ( X k ) / dt = - A Xk + ( 1 - B Xk ) Jk + - ( C + D Xk ) J k (2.1)
Where, Xk = activation function of the kth unit,Jk + =
excitatory input to the kth unit ,Jk- = inhibitory input to the kth
unit ,Under steady state (asymptotic) conditions (using B = C
= 0 for ART 2) , this equation reduces to:
X k = J k + / ( A + D J k -)
(2.2)
Activation functions of each sub layer of F 1 and the orienting
subsystem are calculated using Table 5.1 followed by the
Page 2 of 6
interconnections of the nodes, shown in Fig. 1.1 .The
function f(x) is a noise filter which is described as
f(x) = 0
for 0 ≤ x ≤ Ө
( 2.3)
= x
for
x ≥Ө
where ‘ Ө ‘ is the noise factor which determines the degree of
suppression of noise.Bottom-up (F1 to F2) equation is given
by, T j = ∑ Pi Z ji
(2.4)
A contrast enhancement is used to calculate the activation
output g(yj) of the F 2 nodes g(yj)= d
for T j = Max (T k )
for all k = 0
otherwise
(2.5)
Tk does not include the node recently reset by the orienting
system and d is a constant determined experimentally. The
weights z ij and z ji, corresponding to the interconnections
between y and p nodes form the LTM of the system are
updated as follows zij = zji = u / (1-d)
(2.6)
Only the weights corresponding to the winning node in F2
layer are updated as designated by the J. G node is a gain
control unit and is in-charge of sending an inhibitory signal to
each unit on the layer it is connected. This gain control signal
simulates the ‘2/3’ rule used in ART 1 architecture.
III. STM and LTM Equations and
Synaptic Plasticity
The STM activity xk of any node vk in F1 or F2 obeys a
membrane equation of the form
 d(xk) /dt= -xk + (1-Axk)Jk+ - (B+Cxk)Jk-
(3.1)
Jk+
Where
is the total excitatory input to vk . Jk- is the total
inhibitory input to vk and all the parameters are non negative.
If A>0 and C>0, then the STM activity xk(t) remains within
finite interval [-BC-1 to A-1] no matter even for large values
of nonnegative inputs J k+ and Jk-..We denote nodes in F1 by vi
where i =1,2…M. We denote nodes in F2 by vj where
j= M+1, M+2…..N. Thus,
 d(xi) / dt = -xi + (1-A1 xi)Ji+ - (B1 +C1 xi)Ji- -- (3.2)
 d(xj)/dt= -xj + (1-A2 xj)Jj+ - (B2 +C2 xj)Jj-
(3.3)
The notation is, the F1 activity pattern x = (x1, x2,…., xM) and
the F2 activity pattern
y = (x M+1, xM+2,.., xN).The
input Ji+ to the ith node vi of F1 is a sum of the bottom-up input
Ii and the top-down template input vi
vi = D1  f(xj) zji
And
Ji+
(3.4)
= Ii + vi
Where f (xj) is the signal generated by activity xj of vj and zji is
the LTM trace in the top-down pathway from vj to vi. In the
notation, the input pattern I=(I1, I2, …, IM), the signal pattern
u=(f (xM+1), f (xM+2), , f(xM)), and the template pattern v=(v1,
v2,…, vM). output patterns are described by S=(h(x1), h(x2), ,
h(xM)) and input pattern T= (T M+1, TM+2, ,TM). In this study the
model system is recurrent networks of excitatory and
inhibitory type of integrate and fire neurons. Excitatory
neurons are connected by plastic synapses having two values
of
potentiated and depressed efficacy (PSP), which are
assumed to be stable on the long time scale. Long term
synaptic dynamics, i.e transitions between these two states is
driven by neural pre synaptic and post synaptic activities
through a fast, spike driven dynamics which affects the
internal variability of the synapse. Up or down regulation of
the internal variable occurs if a pre-synaptic spike comes
within (beyond) t. When the internal variable crosses a
threshold from below, the efficacy gets potentiated or
depressed. Because of the stochasticity of the spikes in the
network (due to features like the random pattern of
connectivity) such synaptic transitions are themselves
stochastic, and are characterized by a frequency dependent
‘transition probability’. Such a synaptic device, driven by
spike timings, acts as pre- and post- synaptic neurons, and
implements a rate dependent memory. For a network with N
neurons the analysis of the coupled neural synaptic dynamics
requires to solve a system of O (N+N2) coupled equations.
Standard approaches to the numerical integration of those
equations imply a prohibitive O (N3) computational load.We
therefore use the event driven approach to stimulate the
complexity of O(N2). On the other hand, predictive, analytical
tools are needed in order to choose interesting regions in the
huge parameters space of the neural and synaptic dynamics.
As long as the characteristic times of synaptic changes are
much longer than those neural dynamics, it still serves as a
precious tool in the present scenario to analyze the working
memory. Assuming that synapses undergoing potentiation and
depresses homosynaptically, a compact representation can be
deviced, exposing in the plane (JP and JD ) of the stable
collective states predicted by MFT . We will be seeking
learning trajectories bringing the network from a totally
unstructured synaptic state (the lower right corner in the plane)
to a state supporting a low rate. An unspecific “spontaneous”
state coexisting with a selective higher rate state is interpreted
as neural substrate of a “working memory”. To take STD into
account, we adopt the model introduced in [7,8] and calculate
the corrections to the mean and variance of the apparent
currents (the needed ingredients for MFT) due to inclusion of
STD [9].
The role of LTD and STD with feasibility, it is clear that
despite the wide region allowing for coexistent spontaneous
and selective activity, its learning trajectories have to cross
from below, has a very steep rise with respect to J p (Jp is
candidate source of instability). If the increase of J p is not
adequately balanced by increased depression, the spontaneous
activity is easily destabilized. The stabilizing role of the LTD
is such that the lower Jp /Jee (higher depression) the wider the
regions of the synaptic space in which the spontaneous and
selective activities coexist, and the region of instability
shrinks. The inclusion of STD, the spontaneous activity is
virtually unaffected and is kept stable in much wider region of
the plane while the rates of the selective states have a very
smooth dependences on Jp,
suggesting that, simulations
confirm on line learning, the ongoing neural synaptic
dynamics(diamonds in the plot) meet the above expectations.
Keeping spontaneous activity stable during learning implies
that synapses driven by low neural activity should not change.
Page 3 of 6
During learning a sufficiently intense synaptic long term
depression should take place, in order to compensate for
synaptic potentiation [10].
A multilayer ART 2 model with
varying number of input nodes and output nodes using neural
network simulator running with MAT Lab and GENESIS
simulator is used. During training small amount of noise is
added to the original training input to create large number of
slightly different patterns. During the ID thousands of the inter
neurons synchronously undergo an unusual large
depolarization on which burst of action potential is
superimposed and neuronal inhibition occurs in areas
surrounding the focus [11,12].
IV. Training and Testing
Procedure for ART 2 Network:
Since the ART2 network is essentially an unsupervised
learning system, the output clusters have to be labelled using
some kind of an external event or supervision. Such labeling
can simply be done by a trainer who assigns desired classes to
the different clusters as shown in Fig. 4.1. In the training
phase, the system trainer assigns class names (spike vs. nonspike) to clusters and a node classification table is generated.
In the testing phase, this table is used to classify the node
generated by ART2. New nodes generated by ART2 during
the testing phase are labeled as unclassified nodes.
are used for training the network. Each EEG waveform was
included in only one set in order to ensure that testing was
performed on new waveforms not used during learning. All
these networks were trained by the sets of EEG data and tested
with all other standard three sets of data. Percentage correct
classification over the whole test sets gave the overall success
of the network in comparison to expert ART network. In order
to evaluate the potential of the unsupervised ART2 network,
the same data were evaluated by using the standard back
propagation (BP) neural network. This BP network consisted
of 32, 8 and 2 nodes in the input, hidden and output layers.
V. Results
The clustering of a set of EEG data containing 30 spike
and 30 non-spike patterns is used for training the network. As
seen, the network is iterated to produce 8 nodes for spikes and
15 nodes for non-spikes. As expected more nodes were
generated for non-spikes since they can be of any shape
observed in EEG. Surprisingly it was also observed certain
seizure waveforms at the output nodes of the constructed
ART2 module iterated with long term synaptic memory as
illustrated in Fig. 5.2 to Fig. 5.3. The number of nodes and the
accuracy of the network is strongly influenced by the vigilance
factor ( ρ ) and to a lesser degree by the noise factor ( Ө ). In
this study, we have chosen a value for the vigilance factor by
experimentally incrementing it in small steps and scanning
through an estimated range. The noise factor is recommended
to be set at l/√n where n is the number of input nodes.
Generally, for higher Ө, more nodes are generated and
accuracy goes down. On the other hand, for lower Ө, fewer
nodes are generated but accuracy goes down. Also, the
network becomes unstable. We have determined Ө = 0.223.
The overall results comparing the supervised network (BPNN) with the unsupervised network (ART2-NN) are shown in
Table 5.1 and 5.2. ART 2 neural network generally produce
lower accuracies since these networks are not trained under
supervision. However, the ART 2-NN results were very close
to the performances of the BP-NN nets indicating the high
future potential of this approach.
Fig.4. 1. Training and testing procedures using
unsupervised neural architecture.
The ART neural network system described above is
implemented. EEG data is recorded from epileptic patients
undergoing seizure treatment were used for training and
testing. Digitized temporal sequences (32 points) of EEG data
were organized in three sets. Each set consists of an equal
number of spike and non-spike waveforms as determined by
the standard EEG data.
These sets consists of 60 and 150 EEG waveforms
obtained from 22,24 and 26 patients,
(ASC 60/22,
ASA150/24, ARA 150/26 are named as data files) respectively
Page 4 of 6
and accuracy level achieved are clinically acceptable and
comparable to those of human oral observations.
Fig. 5.2 Detection of spikes(S) and artifacts (A)
Fig. 5.3 Linear combination for spike and wave detection
1
Fig. 5.1 Spike and non-spike wave patterns generated by
ART 2 at its output nodes.
Table: 5.1 Spike and wave detection
VI. Discussions and Conclusion
It has been concluded that the ART 2 neural network
model reproduces many of the observed activities of
epileptical seizure. Epileptical seizure data can be recalled and
its characterization can be studied using ART 2 networks. We
have outlined a systematic methodology, which reduces the
construction of an iterative neural network, which closely
models time series data from a given chaotic system, to a
relatively automatic process. The dependence of the number of
nodes and classification accuracy on the vigilance factor and
threshold noise gives an opportunity for the user to optimize
the network by fine tuning these parameters during online
testing and thus ART 2 module can itself act as a memory to
generate certain observed behaviors of epileptic disorders. The
added stability and plasticity feature of ART 2 networks, and
the speed at which network converges has proven that, this
method of constructing working memory to characterize
epileptical EEG seizure is best compared to other conventional
neural network algorithms.
References:
Table 3.2 Performance comparisons of supervised BP and
unsupervised ART 2 NN.architectures using the training
and testing files.
Performance is measured in percentage. Testing of BPNN with ASC and ASA gave about 90% accuracy. This
performance, however, dropped to about 80% when an
inflexible set of data (ARA) containing waveforms from
many new patients was tested. As previously described,
performance of the networks are generally lower when they
are subjected to totally new waveforms from new subjects.
Overall, the test results indicate that it is possible to automate
spike and seizure detection of epilepsy with ART2 network
1. Victor, J. D., and Canel, A., “A Relation Between the Akaike Criterion and
Reliability of Parameter Estimates, with Application to Nonlinear
Autoregressive Modelling of Ictal EEG”, Annals of Biomedical Engineering,
vol. 20, pp. 167-180, pp. 167-180, 2011
2. Wang J., and Wieser, H. G., “ Regional Rigidity of Background EEG
Activity in the Epileptogenic Zone, Epilepsia”, vol. 35 (3), 2004, pp. 495-504.
3. Webber, W. R. S., Litt B., Wilson K., and Lesser R. P., “ Practical
Detection of Epileptiform Discharge (ED's) in the EEG using an Artificial
Neural Network: A Comparison of Raw and Parameterized EEG data”,
Electroencephalography and Clinical Neurophysiology, vol. 91, pp. 194-204,
2014.
4. Wendling F., Shamsollahi, J. M., Bellanger J. J., “ Time-Frequency
Matching of Warped Depth-EEG Seizure Observations”, IEEE Transactions
on Biomedical Engineering, vol. 46, no. 5, pp. 601-605, May 1999.
5. Werman M., and Peleg, S., “Min-max Operators in Texture Analysis”,
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 7, pp.
730-733, 2010.
Page 5 of 6
6. Williams, W. J., Zaveri, H. P., and Sackellares, J. C., “ Time-Frequency
Analysis of Electrophysiology Signals in Epilepsy”, IEEE Engineering in
Medicine and Biology, Mar/April, 2009, pp. 133-143.
7. Witte, H., Eiselt, M., Patakova, I., Petranek, S., Griessbach, G., Krajca, V.,
and Rother, M., “ Use of Discrete Hilbert Transformaion for Automatic Spike
Mapping: a Methodological Investigation”, Medical & Biological Engineering
& Computing, vol. 29, pp. 242-248, 2010.
8. A.Barreto, J.Andrian, N.Chin and J.Riley (2011). “Multiresolution
characterization of interictal spikes based on a wavelet transformation”.
Wellesley-Cambridge Press, Wellesley MA.
9.. A.Graps (1995). “ An introduction to wavelets” IEEE Computational
Science and Engineering,Los Alamitos, USA.
10.E.Niedermeyer and F.L.Silva (2008). “Electroencephalography:Basic
principles, clinical applications and related fields”. Williams and
Wilkins,Baltimore, Maryland, USA.
11. G.Strang, T.Nguyen (2009). “Wavelets and Filter Banks”. WellesleyCambridge Press, Wellesley.
12. J.M.H.Karel, R.L.M.Peeters and R.L.Westra (2005). “Optimal discrete
wavelet design for cardiac signal processing”. accepted, EMBC 2013.
First Author P.Aswathappa :
At present he is Working as a Research scholar, Dept. Of ECE,
Karpagam University,coimbotore,T.N,, India. Also he is working as
a Assistant Professor in the Dept. of IT, Bangalore Institute of
Technology, Bangalore Affiliated to Visveswaraya Technical
University. he is good exposure in the field of signal
processing, ,Image processing, etc. he got more 16 years of
Teaching experience in various field.
Dr. H.N. Suresh received his BE
(E&C) from P.E.S College of
Engineering, Mysore University,
Karnataka, India, in the year 1989
and completed his M.Tech (Bio
Medical Instrumentation) from
SJCE
Mysore
affiliated
to
University of Mysore., in the year
of 1996 and since then he is actively involved in teaching and
research and has Twenty six years of experience in teaching.
He obtained his PhD (ECE) from Anna university of
Technology.He worked at various capacities in affiliated
University Engineering Colleges. For Visveswaraya Technical
University and Bangalore University he worked as a Chairman
for Board of Examiners, and member of Board of Studies etc.
At present he is working as Professor and Research and PG
co-ordinator in Bangalore Institute of Technology, Bangalore
,Affiliated to Visveswaraya Technical University. He has good
exposure in the field of signal processing, Wavelet
Transforms, Neural Networks,Pattern recognition,Bio Medical
Signal Processing, Netwoking and Adaptive Neural network
systems. He has published more 45 research papers in the
refereed international journals and presented contributed
research papers in refereed international and national
conferences. He is a member of IEEE, Bio Medical Society of
India, ISTE,IMAPS & Fellow member of IETE.
Page 6 of 6