Comprehensive Analysis of Swarm Based Classifiers and Bayesian Based Models for Epilepsy Risk Level Classification from EEG Signals
©2017
Textbook
48 Pages
Summary
This project presents the performance analysis of Particle swarm optimization (PSO), hybrid PSO and Bayesian classifier to calculate the epileptic risk level from electroencephalogram (EEG) inputs. PSO is an optimization technique which is initialized with a population of random solutions and searches for optima by updating generations. PSO is initialized with a group of random particles (solutions) and then searches for optima by updating generations. Hybrid PSO differs from ordinary PSO by calculating inertia weight to avoid the local minima problem. Bayesian classifier works on the principle of Bayes’ rule in which it is the probability based theorem.
The results of PSO, hybrid PSO and Bayesian classifier are calculated and their performance is analyzed using performance index, quality value, cost function and classification rate in calculating the epileptic risk level from EEG.
The results of PSO, hybrid PSO and Bayesian classifier are calculated and their performance is analyzed using performance index, quality value, cost function and classification rate in calculating the epileptic risk level from EEG.
Excerpt
Table Of Contents
4
LIST OF TABLES
Table 2.1 Average Values of Extracted Parameters from Patient Record 4 ... 18
Table 2.2 Representation of Risk Level Classifications ... 21
Table 4.1 Performance Index of classifiers of Wavelet Transform along
hard Thresholding ... 31
Table 4.2 Performance Index of classifiers of Wavelet Transform along Soft
Thresholding ... 31
Table 4.3 Quality Value of classifiers of Wavelet Transform along hard
Thresholding ... 32
Table 4.4 Quality Value of classifiers of Wavelet Transform along Soft
Thresholding ... 33
Table 4.5 Time Delay of classifiers of Wavelet Transform along hard
Thresholding ... 33
Table 4.6 Time Delay of classifiers of Wavelet Transform along Soft
Thresholding ... 33
Table 4.7 Mean Square Error of classifiers of Wavelet Transform along
hard Thresholding ... 34
Table 4.8 Mean Square Error of classifiers of Wavelet Transform along
Soft Thresholding ... 34
Table 4.9 Missed classification and False alarm of classifiers of Wavelet
Transform along hard Thresholding ... 35
Table 4.10 Missed classification and False alarm of classifiers of Wavelet
Transform along soft Thresholding ... 35
Table 4.11 Perfect classification of classifiers of Wavelet Transform along
hard Thresholding ... 36
Table 4.12 Perfect Classification of classifiers of Wavelet Transform along
soft Thresholding ... 36
Table 4.13 Overall Performance of PSO, Hybrid PSO and Bayesian classifier . 38
5
LIST OF FIGURES
Fig 1.1 EEG Recording by 10-20 system ... 12
Fig 1.2 Sample 2-second epoch ... 14
Fig 2.1 Classification Overview ... 17
Fig 4.1 Measure of performance Index of three classifiers with wavelet
thresholding ... 39
Fig 4.2 Measure of Mean Square Error of three classifiers with wavelet
thresholding ... 39
Fig 4.3 Comparison of performance of the three classifiers with wavelet
thresholding ... 40
Fig 4.4 Measure of Quality Value of three classifiers with wavelet
thresholding ... 40
Fig 4.5 Measure of Perfect classification of three classifiers with wavelet
thresholding ... 41
Fig 4.6 Measure of Missed Classification of three classifiers with wavelet
thresholding ... 41
Fig 4.7 Overall performance of three classifiers with wavelet thresholding ... 42
6
LIST OF ABBREVIATIONS
PSO
Particle Swarm Optimization
EEG
Electroencephalogram
EEGer
Electroencephalographer
CBF
Cerebral Blood Flow
DWT
Discrete Wavelet Transforms
PC
Perfect Classification
MC
Missed Classification
FA
False Alarm
7
CHAPTER 1
INTRODUCTION
The motto of this project is to design and simulate a classifier using PSO, hybrid PSO and
Bayesian classifier to classify the epilepsy risk level of the given input EEG signal.
Classification is a basic task in data analysis and pattern recognition that requires the
construction of a classifier, that is, a function that assigns a class label to instances described
by a set of attributes.
The use of EEG signals as a vector of communication between men and
machines represents one of the current challenges in signal theory research. The principal
element of such a communication system is brain computer Interface which is the
interpretation of the EEG signals related to the characteristic parameters of brain electrical
activity. The Cerebral Blood Flow (CBF) measurements, Electroencephalogram (EEG) signals
are the input parameters, sudden, recurrent and transient disturbances of brain functions or
movements of body that results from excessive discharging of groups of brain cells
characterize epilepsy. In clinical neurological practice, detection of abnormal EEG activity
plays an important role in diagnosis of epilepsy.
It is often difficult to identify and model the likelihood of epilepsy risk level through
traditional modeling tools or techniques. The dichotomous nature of conventional logic is
inadequate in representing the stages that an individual may undergo in the transition from
the condition of normal to the condition of high epilepsy risk level [6]. Conversely, the multi-
valued property of Particle swarm optimization technique allows it to be a useful tool in the
representation of different epilepsy risk levels as it develops [2]. Likewise classsifier is
particularly useful when considering epilepsy risk level because this may develop over a
period of weeks, months or years. The above mentioned properties of PSO allow for the
evaluation of the gray area in the condition of epilepsy selected as the decision.
1.1 GENESIS OF EEG SIGNALS
The word Electroencephalography (EEG) was derived from the Greek words "electro"
`enkephalos' and `graphy'. Therefore, the literal translation of EEG would be the writing, or
study of the electrical signals in the brain. Electroencephalograph would record the electrical
activity taken from the human scalp over, usually, a period of time. The sensors would be
placed in multiple locations of the subject's scalp. Recordings of the electrical signals are
simultaneously performed for all channels. From the signal processing viewpoint, EEG is a
8
spatial and non-stationary time-series process. An analysis of the EEG signals is one of the
key areas of biomedical data processing due to the information contained in these signals.
EEG could be extremely beneficial for studying the conditions and status of the human brain.
Similar to other naturally-generated signals, EEG also contains information that could be
extracted by using signal processing techniques. Various types of such techniques have been
developed to analyze EEG signals. An accurate analysis could provide valuable clinical,
psychological, and physical information in reference to the brain. In particular, EEG
waveforms would disclose information about certain changes.
The human Electroencephalogram (EEG) is usually recorded from electrodes attached
to the scalp using high amplifiers, which are usually coupled to the scalp electrodes. The
amplified signals are written out on paper via a polygraph, which contains typically 8 to 16
channels. Normal subjects usually exhibit alpha, beta, theta and delta activities, while
abnormal activity may be manifested by a slowing and decrease in amplitude of EEG,
increase in the EEG frequency, and the presence of sudden EEG discharges (paroxysmal
activity) different from the background either in frequency content or amplitude or pattern.
The EEG is a powerful tool for the diagnosis of neurological disorders. Since its discovery,
the EEG has been used for the diagnosis of epilepsy, for trauma assessment, for sleep
research, and for the analysis of higher brain functions. The EEG is highly dependent upon
the availability of high quality instrumentation, and almost from the beginning, automated
methods of signal qualification have been applied. One of the primary goals is to help the
encephalographer (EEGer) in the time consuming task of quantifying signal that appears to
the eye as a low information content background intermixed with either bursts of rhythmic
activity with different frequencies (the EEG rhythms) or short transients of clinical
significance. In spite of years of research to produce universal automated detection methods,
success has been achieved only in specific areas. Accomplishments include automatically
sleep staging with a high degree of accuracy; counting spikes and wave complexes, and
monitoring in intensive care units. However clinicians still rely on visual analysis for clinical
applications.
1.2 EPILEPSY DETECTION AND EEG SIGNALS
Epilepsy is a brain disorder in which clusters of nerve cells or neurons in the brain sometimes
function abnormally. Epilepsy is a neurological condition that makes peoples susceptible to
seizures. A seizure is a change in sensation, awareness, or behavior brought about by a brief
electrical disturbance in the brain. There are many different types of seizures: including ones
9
affecting the whole brain and ones only impacting a part of the brain. The term "seizure" is
widely used to describe an abnormal spasm or convolution, generated by excessive electrical
activity in the brain. In Epilepsy the normal pattern of neuronal activity becomes disturbed,
causing strange sensations, emotions and behavior or sometimes convulsions, muscle spasms
and loss of consciousness. It may develop due to
(i) Abnormality in brain firing,
(ii) An imbalance of nerve signaling chemicals.
Epileptic seizures result from a temporary electrical disturbance of the brain. Sometimes
seizures may go unnoticed, depending on their presentation, and sometimes may be confused
with other events, such as a stroke, which can also cause falls or migraines. Approximately
one in every 100 persons will experience a seizure at some time in their life. Unfortunately,
the occurrence of an epileptic seizure seems unpredictable and its process is very little
understood. Since its discovery by R.Caton, the Electroencephalogram (EEG) has been the
most utilized signal to clinically assess brain activities. Twenty five percent of the world's
50 million people with epilepsy have seizures that cannot be controlled by any available
treatment. The need for new therapies, and success of similar devices to treat cardiac
arrhythmias, has spawned an explosion of research into algorithms for use in implantable
therapeutic devices for epilepsy. Most of these algorithms focus on either detecting
unequivocal EEG onset of seizures or on quantitative methods for predicting seizures in the
state space, time, or frequency domains that may be difficult to relate to the Neuro
physiology of epilepsy. Between seizures, the EEG of a patient with epilepsy may be
characterized by occasional epileptic form transients-spikes and sharp waves. EEG patterns
have shown to be modified by a wide range of variables including biochemical, metabolic,
circulatory, hormonal, neuro electric and behavioral factors.
Exploring various analytical approaches, both linear and non linear methods to
process data from medical database is meaningful before deciding on the tool that will be
most useful, accurate, and relevant for practitioners. For example, assigning a new patient to
a particular outcome class is a classification problem commonly described as "pattern
recognition", "discriminant analysis", and "supervised learning". In the past, the
Encephalographer, by visual inspection was able to qualitatively distinguish normal EEG
activity from localized or generalized abnormalities contained within relatively long EEG
records.
10
Electroencephalography is a well-established clinical procedure, which can provide
information pertinent to the diagnosis of a number of brain disorders (e.g., epilepsy or brain
tumors). However, despite its widespread use, it is one of the last routine clinical procedures
to be fully automated. Analysis of the electroencephalogram (EEG) includes the detection of
patterns and features characteristic of abnormal conditions. For example, Asymmetries in the
amplitude or frequency of background activity suggest a lesion, while the presence of
epileptiform activity supports a clinical diagnosis of epilepsy. Over half the EEG referrals
relate to epilepsy, with the EEG being the most useful procedure in its diagnosis. Recording
the EEG during a seizure is particularly helpful in determining whether a patient has epilepsy.
Because seizures usually occur infrequently and unpredictably, obtaining such recording
might require an EEG extending over several days (long-term EEG monitoring). Techniques
have been developed for the automated detection of petitmal seizures and grand mal seizures,
which have proven relatively successful.
Between seizures, the EEG of a patient with epilepsy may be characterized by
occasional epileptiform transients (spikes and sharp waves) and, consequently, relatively
short recording can still be useful in the diagnosis of epilepsy. A routine recording typically
lasts 20-30 minutes, during which some 4 minutes of paper record are produced[. An
electroencephalographer (EEGer) detects epileptiform transients by visual inspection of the
recording, which requires considerable skill and is time consuming. Hence, automation of
this process could save time increase objectivity and uniformity, and enable quantification for
research studies. Automated detection of epileptiform transients has two primary areas of
clinical application. The first is in long term EEG monitoring, where it acts essentially as a
daily reduction process. A segment of EEG is recorded only when a transient is detected and
all segments are reviewed by an EEGer. Thus, the goal is to detect a high proportion of
epileptiform activity while minimizing false detection. The second area is in routine clinical
recordings where, major objective is to minimize the visual inspection process as far as
epileptiform transients are concerned. In this case it is important not to precipitate a
misdiagnosis of epilepsy and, therefore, the aim is to eliminate false detections while
detecting a satisfactory proportion of epileptiform transients.
Spikes and sharp waves are defined as transients clearly distinguished from
background activity with pointed peaks at conventional paper speeds. Spikes are defined
having durations of 20-70 ms, while sharp waves have durations of 70-200 ms. No distinction
is made between spikes and sharp waves and, therefore, they are collectively termed
11
epileptiform transients. Due to the variety of morphologies of epileptiform transients and
similarities to waves which are part of the background activities and due to artifacts (i.e.,
extra cerebral potentials from muscles, eyes, heart, electrodes, etc.), the detection of
epileptiform activity in the EEG is far from straightforward. Several techniques have been
applied to the detection of epileptiform activity in the EEG. These include:
· Template matching, where detection is made whenever the value relation of the EEG
with a template exceeds a threshold.
· Parametric methods, where a detection is made when the difference between the EEG
and its predicted value used on the assumption that the background is stationary
exceeds a threshold
· Mimetic methods, where one or more parameters of each wave are calculated and
threshold.
· Syntactic methods, where detections are based on the presence of a structural
combination of structures
· Artificial neural networks trained to detect epileptic waveform transients and
· Expert systems, which detect epileptiform activity by mimicking the knowledge and
reasoning of the EEGer.
Most of these systems are in the developmental stage, and those in clinical use are restricted
to long-term EEG monitoring with all detections being reviewed by an EEGer. Due to a high
number of false detections these systems cannot perform satisfactorily in the routine EEG
setting.It is generally accepted that the only way to separate epileptiform from non-
epileptiform waves is to make use of a spatial and temporal context. Several groups are
implementing this approach in an effort to minimize false detections. Glover et al. have
developed a system that relies on a wide spatial context, with 12 EEG channels being
analyzed together with additional contextual information provided by EKG, EOG, and EMG
signals. This system is proven to be particularly successful at rejecting non-epileptiform
activity during awake and resting EEG's. It uses a mimetic the method to detect candidate
transients, which are subsequently trimmed or rejected as epileptiform by an expert system.
The system integrates both spatial and temporal contextual information to detect definite and
probable epileptiform activities and reject non-epileptiform waves. Preliminary results state
that this system should be capable of performing routine clinical EEG setting.
12
1.3 EPILEPSY CLASSIFICATION SYSTEM
The block diagram of fuzzy based epilepsy risk level classifier is shown in figure1.1. This is
accomplished as:
1. Classification of epilepsy risk level at each channel from EEG signals and using PSO,
hybrid PSO optimization and Bayesian neural network technique.
2. Each channel results are optimized, since they are at different risk levels.
The Electroencephalogram signals from epileptic patients are to be collected from hospitals.
Then the EEG signals are then converted to code patterns by fuzzy systems. The output of a
fuzzy system represents a wide space of risk levels. This is due to sixteen different channels
of input to the system in three epochs. This yields a total of forty-eight input output pairs.
Since the known cases of epileptic patients are detected, then it is indispensable to find the
exact level of risk the patient. PSO optimization will also aid in the development of
automated systems that can precisely classify the risk level of the epileptic patient under
observation. Hence an optimization of the outputs of the fuzzy system is initiated. This will
improvise the classification of the patient's state and can provide the EEGer with a clear
picture.
1.4 DATA COLLECTION
The EEG is recorded by placing electrodes on the scalp according to the International 10-20
system. Sixteen channels of EEG are recorded simultaneously for both referential montages,
where all electrodes are referenced to a common potential like ear, and bipolar montages,
where each electrode is referenced to an adjacent electrode. The EEG recording points on the
scalp are illustrated in figure 1.2. Recordings are made while the patient is awake but resting
and include periods of eyes open, eyes closed, hyperventilation and photoic stimulation.
Amplification is provided by an EEG machine (Siemens Minograph Universal).
Fig 1.1 EEG Recording by 10-20 system
13
Before placing the electrodes, the scalp is cleaned, lightly abraded and electrode paste is
applied between the electrode and the skin. By means of this application of electrode paste,
the contact impedance is less than 10 k
. Generally disk like surface electrodes are used. In
some cases, needle electrodes are used to pick up the EEG signals. The signals are recorded
with the speed of 30 mm/s. The obtained signals are filtered by notch filter (low pass filter -
5Hz, high pass filter - 75Hz).The EEG is broken down into sections or epochs, for the
purpose of feature extraction. An epoch of 2.0 s is used for the following reasons:
1) It is long enough to capture the main statistical characteristics of the EEG and short
enough to capture the evolution of seizures
2) The EEG being digitized at a sampling rate of 200 Hz an epoch of 2s contains 400
samples, which is a convenient length for computation.
The software for analyzing the EEG data was implemented using C++ programming and Mat
lab 7.2. Waveforms of normal and abnormal data are plotted and studied.A group of twenty
patients with known clinical findings of epileptic seizure is undertaken for classifications of
level of epilepsy risk.
1.5 FEATURE EXTRACTION
The pixels of the bmp files are converted to x and y coordinates where the y coordinate
represents the signal amplitude value. The signals are reconstructed with the following
scaling factor:
X-axis: 60mm = 2seconds
Y-axis: 1mm = 70
V
The X-axis of the scaled image is set to a width of 400 pixels so that each pixel represents a
sampled amplitude value. These amplitude values are found using graphics programming in
C++ and are written to a file. The features of the epoch which are used for GA optimization
viz., energy, variance, peaks, sharps and spikes, events, average duration, covariance of
duration are computed based on the sampled amplitude values. A two-second epoch of a
single channel is shown in figure 1.3 for which the aforesaid parameters were obtained as
Energy: 2864
Variance:7.156396
Peaks:1
Average amplitude:0.2356 Sharps and Spikes: 33 Events:24
Duration: 0.039035 Covariance of Duration: 0.457347
Details
- Pages
- Type of Edition
- Erstausgabe
- Year
- 2017
- ISBN (PDF)
- 9783960676225
- ISBN (Softcover)
- 9783960671220
- File size
- 839 KB
- Language
- English
- Institution / College
- Bannari Amman Institute of Technology – Department of ECE
- Publication date
- 2017 (February)
- Keywords
- Particle swarm optimization PSO EEG Bayes Electroencephalogram Hybrid PSO Epileptic risk Epilepsy Bayesian classifier
