Research Proposal on

A new offline and online classifier of extracted EEG signal
Abstract:
Seizure is a general term that refers to a sudden malfunction in the brain that causes someone to
collapse, convulse, or have another temporary disturbance of normal brain function, often with a loss
or change in consciousness. Epilepsy is a chronic disorder of the central nervous system that
predisposes individuals to experiencing recurrent seizures. Since 1970 numerous techniques have
been applied for detecting epileptic seizure activities. This proposal is made to introduce a new
method intended to ease the burden of intractable seizures. The method uses machine learning to
construct patient-specific classifiers that are capable of rapid, sensitive, and specific detection of
seizure activities. The classification method uses a set of features or parameters to characterize each
problem and finally draws the conclusion with a specific anomalous history (if any).
Introduction:
The human brain is obviously a complex system and exhibits rich spatiotemporal dynamics. Among
the noninvasive techniques for probing human brain dynamics, electroencephalography (EEG)
provides a direct measure of cortical activity with millisecond temporal resolution. EEG is a record of
the electrical potentials generated by the cerebral cortex nerve cells. There are two different types of
EEG depending on where the signal is taken in the head: scalp or intracranial. There are electrodes
which are placed on the scalp or implanted in the brain by which the voltage of the neuron current is
measured. Then the voltage difference between electrodes is amplified before being transmitted to a
computer program to display the tracing of the voltage potential recordings. Epileptic seizure is an
abnormality in EEG recordings and is characterized by brief and episodic neuronal synchronous
discharges with dramatically increased amplitude. This anomalous synchrony may occur in the brain
locally (partial seizures), which is seen only in a few channels of the EEG signal, or involving the
whole brain (generalized seizures), which is seen in every channel of the EEG signal. EEG signals
involve a great deal of information about the function of the brain. But classification and evaluation
of these signals are limited. Since there is no definite criterion evaluated by the experts, visual
analysis of EEG signals in time domain may be insufficient. Therefore, some automation and
computer techniques have been used for analyzing the signal characteristics.
Method:
The EEG data will be taken from 24h EEG recordings from both epileptic patients and normal
subjects. The patient and subjects will be consisted of different aged people. For analysis purposes
four/five bipolar channels will be taken as reference. Recording will be done under video control to
have an accurate determination of the different stages of the seizures. The data from recordings will
be passed through a band pass filter (1-70Hz). The filtered EEG signal will be segmented to 6s (no of
samples) durations. The EEG found from the specified channels will be digitized for processing.
Process Used to EEG signals:
In the earlier days of automatic, EEG processing included Fourier transform which dictated some
observations that the EEG spectrum contains some characteristic waveforms that fall primarily within
four frequency bands—–delta (<4 Hz), theta (4—8 Hz), alpha (8—13 Hz) and beta (13—30 Hz).
Though it has proved beneficial for various EEG characterizations, but suffers from large noise
sensitivity. Time-Frequency Analysis was carried out along with STFT and various quadratic TFDs.
Again the major problem inherited with quadratic TFDs was interference. Spectral loss problems are
minimized by Autoregressive (AR). But, since the EEG signals are non-stationary, the AR methods
are not suitable for frequency decomposition. To analyze those signals Wavelet Transforms (WT)
method is more suitable. It is well suited to locating transient events, which may occur during
epileptic seizures. The capability of this mathematical microscope to analyze different scales of
neural rhythms is shown to be a powerful tool for investigating small-scale oscillations of the brain
signals.
Feature Extraction:
This portion is an overview of measures that are commonly used to characterize EEG time series.
These are univariate linear/nonlinear measures and bivariate linear/nonlinear measures. Univariate
linear measures included Statistical moments, Spectral band power analysis, Spectral edge frequency
analysis, Auto correlation and Autoregressive modeling. Univariate nonlinear measures included
Correlation sum, Correlation dimension, Correlation density, Correlation entropy, Marginal
predictability, Dynamical similarity index, State space dissimilarity measures, Surrogate time series
and surrogate correction and Loss of recurrence. Bivariate linear measures included Maximum Linear
cross-correlation and Linear-coherence. Bivariate nonlinear measures included Non-linear
interdependence, Dynamical entrainment and Phase synchronization. Most commonly features which
are tried by the experts while EEG analysis are Correlation entropy, Spectral band power analysis and
Autoregressive modeling. Rest of the features can be taken into account during my research and a
comparison can be drawn after analysis.
Classifications of EEG signals:
There are several methods which are utilized to classify the extracted features as seizure or non-
seizure events. Below is given the short description of each method.
1) Naive Bayes Classifier:
The Naive Bayes classifier is designed for use when features are independent of one another within
each class, but it appears to work well in practice even when that independence assumption is not
valid. It classifies data in two steps followed by training and prediction step.
2) k-NN Classifier:
k-NN classifier is based on learning by analogy. Given an unknown sample, it searches the pattern
space neighbors that are the closest to the unknown sample. Closeness is defined in terms of distance.
The unknown sample is assigned the most common class among its neighbors.
3) Decision Trees:
Decision trees find explicit and understandable rules-like relationships among the input and output
variables using search heuristics. Search heuristics use recursive partitioning algorithms to split the
original data into finer subsets. The algorithm has to find the optimum number of splits and determine
where to partition the data to maximize the information gain. Decision trees are built of nodes,
branches, and leaves that indicate the variables, conditions, and outcomes, respectively.
4) Particle swarm optimization:
In computer science, particle swarm optimization (PSO) is a computational method that optimizes a
problem by iteratively trying to improve a candidate solution with regard to a given measure of
quality. PSO optimizes a problem by having a population of candidate solutions, here dubbed
particles, and moving these particles around in the search-space according to simple mathematical
formulae over the particle's position and velocity. Each particle's movement is influenced by its local
best known position and is also guided toward the best known positions in the search-space, which
are updated as better positions are found by other particles. This is expected to move the swarm
toward the best solutions.
5) Logistic regression:
Logistic regression can be considered a special case of linear regression models. However, the binary
response variable violates normality assumptions of the general regression models. A logistic
regression model specifies that an appropriate function of the fitted probability of the event is a linear
function of the observed values of the available explanatory variable.
6) LLS and LDA:
Linear least squares (LLS) method finds a best fitting linear model that minimizes the mean square
error between the system output and the desired output. Linear discriminant analysis (LDA) uses
hyperplane to find the linear combination of features that best separates two or more classes of
objects or events. Usually within-class, between-class, and mixture scatter matrices are used to
formulate the criterion of searching the hyperplane that the distance between means of classes is
minimized and the interclass variance is maximized.
7) Neural networks:
MLPNN (Multi-Layer Perceptron Neural Network) is widely used in biological phenomena. This
computing system is made up of large number of simple, highly interconnected processing elements.
The architecture of MLPNN may contain two or more layers. Input layer is the first layer which its
number of neurons is equal to the number of selected specific features. Output layer is the last layer
which determines the desired output classes. The intermediate layers (hidden) may be added to
increase the ability of MLPNN mostly useful for nonlinear systems. In the present study, MLPNN
consisted of one input layer, one hidden layer with 21 nodes and one output layer. Back propagation
algorithms are used to train a MLPNN.
8) SVM:
Support vector machine (SVM) uses a hyperplane to identity classes. The hyperplane that maximizes
the margin, i.e., the distance from the nearest training points, is selected by SVM. Maximizing the
margins is known to increase the generalization capabilities. The SVM performs structural risk
minimization and creates a classifier with minimized VC dimension.
Advantages and Disadvantages of different Classifiers:
Classifier Advantage Disadvantage
Neural
Network
They can handle problems with very
many parameters, and they are able to
classify objects well even when the
distribution of objects in the N-
dimensional parameter space is very
complex
They are notoriously slow, especially in
the training phase but also in the
application phase.
K-NN They are easy to implement. They can
also give quite good results if the
features are chosen carefully
The method is also rather slow if the
training set has many examples. The most
serious shortcoming of nearest neighbor
methods is that they are very sensitive to
the presence of irrelevant parameters.
Classifier Advantage Disadvantage
Decision Trees Decision trees are usually much faster
in the construction (training) phase
than neural network methods, and they
also tend to be faster during the
application phase.
They are not as flexible at modeling
parameter space distributions having
complex distributions as either neural
networks or nearest neighbor methods. In
fact, even simple shapes can cause these
methods difficulties.
Naive Bayes
Classifier
An advantage of the naive Bayes
classifier is that it only requires a small
amount of training data to estimate the
parameters (means and variances of the
variables) necessary for classification.
The disadvantage is that the classifier can
model only simple distributions.
Logistic
Regression
It can add explicit interaction and
power terms, It does not require that
the independents be interval and
unbounded
it requires much more data to achieve
stable, meaningful results
SVM Produce very accurate classifiers, less
over fitting, robust to noise
Computationally expensive, thus runs
slow.
Conclusion:
Though it requires huge no of literature surveys in various branches of mathematics and sciences I
had been trying to figure out some of the EEG features and classifiers throughout the proposal. I will
be manipulating all the features and addressing those classifiers during my research period. With my
best knowledge on MATLAB, I hope, I will be able to establish a new methodology on EEG seizure
detection. I must confess that the only picture I have in my mind from my undergraduate years till
now is that of a dedicated researcher in my chosen field of advanced research; being fully aware of
the intellectual rigor and fortitude it claims. I firmly believe in the lesson life taught me over past few
years: our lives are too short to relish our dreams; making reality out of dreams is what does matter. I
just need the right opportunity to prove myself worthwhile.