Real-time Prediction of Epileptic Seizures using EEG...

Real-time Prediction of EpilepticSeizures using EEG signal

A thesis submitted in partial fulfillmentof the requirements for the degree of

BACHELOR OF TECHNOLOGY

in

Electrical Engineering

by

Gorish AggarwalEntry No. 2012EE10454

Under the guidance of

Dr. Tapan K. Gandhi

Department of Electrical Engineering,Indian Institute of Technology, Delhi.

May 2016.

Certificate

This is to certify that the thesis titled Real-time Prediction of Epileptic

Seizures using EEG signal being submitted by Gorish Aggarwal for the

award of Bachelor of Technology in Electrical Engineering is a record

of bona fide work carried out by him under my guidance and supervision at

the Department of Electrical Engineering. The work presented in this

thesis has not been submitted elsewhere either in part or full, for the award

of any other degree or diploma.

Dr. Tapan K. Gandhi

Department of Electrical Engineering

Indian Institute of Technology, Delhi

Abstract

Around 1-2% of the worlds population suffers from epilepsy and 25% of the

epilepsy patients are not curable through surgery or medicine. Epileptic

seizures often occur unpredictably and may cause serious damage to the

patient in some situations, for example, getting a seizure while driving or

crossing a road may lead to a person losing his life. In such a scenario,

a reliable mechanism to predict the onset of seizure beforehand is much

desirable. Past work has been dedicated to the detection of seizure activity

by analysis of EEG data [1] [2]. However, the presence of certain patterns in

the EEG in the pre-ictal stage, indicating a future occurrence of a seizure, has

been established recently [3]. In this study, we shall focus towards reliable

real-time prediction & detection of epileptic seizures.

In this thesis, we initially implemented a seizure detection algorithm using

discrete wavelet transform and Support Vector Machine (SVM) classifier to

detect the presence of seizures in a given EEG signal for a single channel data.

Furthermore, We used the Mean Phase Coherence (MPC) as a measure of

phase synchronization to predict the impending seizures in a multi-channel

EEG data. We showed that during the pre-ictal stages, the MPC values

and thus the synchronization between various channels was found to drop

significantly below the level in an non-ictal EEG signal. Thus, we used

this concept to develop a real-time model in MATLAB for the prediction of

epileptic seizures prior to their actual occurrence. We have also implemented

the above algorithm & technique in Hardware on a Raspberry Pi architecture.

Acknowledgments

I, Gorish Aggarwal, would like to express my deepest gratitude towards my

Bachelor’s Thesis supervisor Dr. Tapan Gandhi, for his invaluable knowl-

edge, motivation and constant support during the project. Tapan sir has

been very supportive during the project. He has provided insightful discus-

sions about the ongoing research on epilepsy. Hw has always guided us to

work towards the betterment of the society and to dream big. Working with

him for my B.Tech project has truly helped me to build a greater under-

standing of this field.

I would also like to thank Mr. Piyush Swami, Mr. Sanjeev Nara and Mr.

Chetan Ralekar for giving us advice, opinions and suggestions for further

improvement in the project. They also gave the much needed data for the

project. The time I spent with the team and at the Neuro-Computing lab

was a great learning experience for me. It helped me to develop my appetite

for this field of research and I would love to continue working in this field

later on in my life so as to help society in a better way. I would also like to

thank Akash Verma, my project partner for the first phase of the project.

We have worked together for the developing the algorithm for the detection

of seizure and giving a proof of concept for seizure prediction in the BTech

Project - 1. His academic genius helped to take the project a long way for-

ward. In the end, I would like to thank my teachers, parents and friends for

their constant support and encouragement.

Gorish Aggarwal

2012EE10454

Contents

1 Introduction 1

2 Literature Review 4

2.1 Previous research on detection of seizure . . . . . . . . . . . . 4

2.2 Previous research on prediction of seizures . . . . . . . . . . . 6

3 Background Work 9

3.1 Electroencephalography . . . . . . . . . . . . . . . . . . . . . 9

3.2 EEG frequency bands . . . . . . . . . . . . . . . . . . . . . . . 10

3.3 Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . 11

3.4 Morphological Operations . . . . . . . . . . . . . . . . . . . . 12

4 Detection 16

4.1 Bonn University Data . . . . . . . . . . . . . . . . . . . . . . 16

4.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.2.1 Pre-processing of Data . . . . . . . . . . . . . . . . . . 17

4.2.2 Decomposition of Data into separate frequency bands . 17

4.2.3 Feature Set Extraction . . . . . . . . . . . . . . . . . . 18

4.2.4 Training the classifier . . . . . . . . . . . . . . . . . . . 19

4.3 Testing & Results . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.3.1 Feature set analysis . . . . . . . . . . . . . . . . . . . . 20

4.3.2 Accuracy, sensitivity, Specificity . . . . . . . . . . . . . 23

4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

c© 2015, Indian Institute of Technology Delhi

5 Prediction 26

5.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.2.1 Mean Phase Coherence . . . . . . . . . . . . . . . . . . 27

5.2.2 Spike Rate . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.2.3 Phase synchronization using bi-variate empirical mode

decomposition(BEMD) . . . . . . . . . . . . . . . . . . 31

5.3 Testing and Results . . . . . . . . . . . . . . . . . . . . . . . . 31

5.3.1 Mean Phase Coherence . . . . . . . . . . . . . . . . . . 31

5.3.2 Spike Rate . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.3.3 Bi-Empirical Mode Decomposition . . . . . . . . . . . 41

5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

6 Real Time Prediction Model 44

6.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

6.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.2.1 Patient with Non-generalized Epilepsy . . . . . . . . . 45

6.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.2.2 Patient with Generalized Epilepsy . . . . . . . . . . . . 49

7 Conclusions and Future Work 51

Bibliography 53

List of Figures

3.1 Brain Electrodes Location [4] . . . . . . . . . . . . . . . . . . 10

3.2 One cycle of Haar wavelet transform . . . . . . . . . . . . . . 12

3.3 Erosion of an image [5] . . . . . . . . . . . . . . . . . . . . . 13

3.4 Dilation of an image [5] . . . . . . . . . . . . . . . . . . . . . 14

3.5 Opening of an image [5] . . . . . . . . . . . . . . . . . . . . . 15

3.6 Closing of an image [5] . . . . . . . . . . . . . . . . . . . . . . 15

4.1 Seizure free signal in the delta band . . . . . . . . . . . . . . . 20

4.2 Seizure signal in the delta band . . . . . . . . . . . . . . . . . 21

4.3 Entropy of samples for their wavelets in the alpha band . . . 22

4.4 Maximum Power spectral density for the samples correspond-

ing to their wavelets in the alpha band . . . . . . . . . . . . . 23

4.5 Kurtosis of data samples corresponding to their wavelets in

the Alpha band . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4.6 Kurtosis of data samples corresponding to their wavelets in

the delta band . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.1 Seizure Activity for Patient 1 . . . . . . . . . . . . . . . . . . 32

5.2 MPC for patient 1 for channels O2 and T6 . . . . . . . . . . . 33

5.3 MPC for patient 1 for channels O2 and P4 . . . . . . . . . . . 34

5.4 MPC for patient 1 for channels F8 and T4 . . . . . . . . . . . 34

5.5 Seizure Window for patient 2 . . . . . . . . . . . . . . . . . . 35

5.6 MPC for patient 2 for channels C4 and P4 . . . . . . . . . . . 36

5.7 MPC for patient 2 for channels C3 and P3 . . . . . . . . . . . 36

5.8 MPC for normal person for channels T4 and F8 . . . . . . . . 37

5.9 MPC for normal person for channels F4 and FP2 . . . . . . . 37


5.10 Seizure Interval of Patient 1. Note that the seizure is present

in left hemisphere only . . . . . . . . . . . . . . . . . . . . . . 39

5.11 Spike Rate in Channel PG1(left hemisphere) . . . . . . . . . . 40

5.12 Spike Rate in Channel PG2(right hemisphere) . . . . . . . . . 40

5.13 Spike Rate in Channel FP2 . . . . . . . . . . . . . . . . . . . 41

5.14 Spike Rate in Channel PG1 . . . . . . . . . . . . . . . . . . . 42

5.15 Plot of Mean phase coherence using BEMD for channels PG2

and F8. The plot corresponds to the phase coherence values

of 3rd IMF (intrinsic mode function) . . . . . . . . . . . . . . 42

6.1 Mean Phase Coherence values between channels T5 and P3.

Blue denotes inter-ictal stage, Green denotes pre-ictal, Red

denotes ictal stage . . . . . . . . . . . . . . . . . . . . . . . . 46

6.2 Mean Phase Coherence values between channels T5 and O1 . . 46

6.3 Mean Phase Coherence values for various channel pairs in left

hemisphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

6.4 Mean Phase Coherence values between channels FP2-F8 (right

hemisphere) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

6.5 Mean Phase Coherence values between channel pair related to

occipital zone . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.6 Mean Phase Coherence values between occipital zone and P4

in non-seizure hemisphere . . . . . . . . . . . . . . . . . . . . 49

6.7 Mean Phase Coherence values between channel pairs(other

than occipital lobe) in non-seizure hemisphere . . . . . . . . . 50

Chapter 1

Introduction

Epilepsy is a chronic brain disorder characterized by recurrent seizures. They

are brief stretches of involuntary movement affecting either a part of the body

or the entire body and occur due to the excessive electrical discharges in a

group of brain cells. Seizures can vary in their origin location, distance of

spread as well as the time and frequency of occurrence.

According to the World Health Organization[6], Epilepsy affects around 50

million people worldwide, making it one of the most common neurological

disease, along with Stroke and Alzheimers disease. Each year, around 2.4

million people are diagnosed with epilepsy globally. Epilepsy is a huge prob-

lem in developing nations where it affects around 40 millions people. The lack

of knowledge combined with non-existence of proper treatment for epilepsy

makes the problem enormous for the people of developing countries. Reports

show that only around 5% of the people diagnosed with epilepsy in India

receive treatment and care for the same[7]. The rest, majority of which are

rural people, do not get benefit of specialized treatment.

There is no known cause associated with epilepsy in most of the cases. How-

ever sometimes, brain injury, stroke, brain tumor and developmental abilities

have been identified to cause epilepsy in some patients. The major symptoms

of epilepsy are continued occurrence of seizure attacks. The younger and the

older generation are known to be most prone to epileptic seizures, though

it can be contracted by anyone at any time. Since the attack themselves

are highly unpredictable in nature and magnitude, they can cause serious

damage to the patient in some situations .

Therefore, a low cost, easily accessible, point of care testing method for the

detection and prediction of seizure is required, to monitor the patient’s men-

tal condition and warn him in advance of an impending seizure attack.

In this project, we have reviewed many of the existing techniques used for

detection and prediction of seizures. During the literature review, we have

come across many linear and non linear techniques for analysis and process-


2

ing of EEG data for the prediction of seizures. We have designed a real time

prediction algorithm which can predict an impending seizure in the near fu-

ture. Such an algorithm would go a long way in building a point of care,

cost effective and economically viable device for the prediction of epileptic

seizures. We are also trying to build a hardware architecture for direct EEG

signal collection and epileptic seizures prediction. Such a device would be a

revolutionary product and a first of its kind to be build. We are planning

to implement the structure on a Beagle board so that Integration of EEG

headset is possible.

The objectives achieved by us during the duration of the project are: BTech

Project Part-1 (Along with Akash Verma):

• Literature Review of the previous existing techniques for detection and

prediction of seizure.

• Classified EEG signals into their ictal and non-ictal stages by using our

seizure detection algorithm and an extensive feature set. Achieved a

good accuracy(94.5%) and high senstivity(96.25%) on the Bonn Uni-

versity Dataset.

• Provided a proof of concept for our seizure prediction algorithm. Demon-

strated that the Mean Phase Coherence values between the electrode

pairs in pre-ictal stages are less as compared to those in inter-ictal

stages.

BTech Project Part-2:

• Extensive testing of the previously developed prediction algorithm on

more than 10 subjects to ensure its correctness and functioning.

• Implemented a real-time MATLAB based on the above concept which

can take EEG signals as inputs in real-time and process each window

sequentially.

• Literature Review of other existing seizure prediction techniques to

increase the pool of features set to be used for classification of pre-ictal

and ictal stages.


3

• Used Spike Rate as another parameter for predicting the onset of an

epileptic seizure. Developed the MATLAB model for finding the ictal

spike rate in an EEG signal.


Chapter 2

Literature Review

Research work on Detection and Prediction of Epileptic Seizure using EEG

has been going on for more than 60 years but only from the mid 1990s has

there been a significant increase in the methods used to detect and predict

the seizures using EEG. The common thread throughout the research is the

deviation of EEG signal from the normal condition during and before the

seizure period (pre-ictal) [8] . With the recent advances in technology and

knowledge, pioneering research groups could now conceivably attempt to de-

velop a system that could detect an ongoing seizure or predict the future

onset of a seizure, thus helping the people prone to seizure to be careful in

advance about an imminent seizure.

Many research groups have been trying to devise various methods of using

EEG data from the brain electrodes to automatically act as an alert system

for the patient to warn him or his caregivers of an impending seizure so that

precautionary measures can be taken in time.[9]. Researchers are also trying

to use the warning system for triggering seizure control mechanisms. Vari-

ous linear and non-linear signal processing techniques coupled with machine

learning algorithms for both uni variate as well as multivariate data have

been used for the above. [10] In this chapter, we shall discuss the major

techniques used for the detection and prediction of seizures in the past have

been discussed and then, we shall give a brief overview into the techniques

applied by us for our work.

2.1 Previous research on detection of seizure

Although the relationship between the mental states and the rhythmic char-

acter of the EEG has been known from a long time, the mechanisms to detect

the mental states(for eg: seizures, sleep or awake condition) of a person from

an EEG data were built only in 1990s. Recently, many studies have also


2.1 Previous research on detection of seizure 5

shown that EEG recording can be used to predict seizures even before the

actual seizure occur. The most common seizure detection techniques uses a

machine learning based classifier to train the feature sets extracted from the

EEG data. We reviewed four major papers which uses this techniques for

the detection of seizures.

Panda et al. [11] used a wavelet transform to decompose signals into different

frequency bands and a SVM classifier to separate epileptic data. The Bonn

university data set [12] was used as a sample for training and testing data

set. The feature set used for the detection purpose included energy, entropy

and standard deviation for each of the frequency bands decomposed using the

Daubechies Discrete Wavelet Transform (DWT-db2). The accuracy achieved

was 91.2% with time for detection around 0.5 sec.

Srinivasan et al. [13] have used approximate entropy measure as a feature

and implemented a neural network based classifier, to detect the presence of

seizure in EEG recording.

Similarly, Ahammad et al. [14] also used the Bonn university data set and

the CHB-MIT scalp database[60] for the detection and onset detection of

seizure respectively. Their feature set included the mean, standard devia-

tion, entropy, power spectral density as well as some statistical features like

inter quartile range and mean absolute deviation. The reported accuracy

was 84.2% with a linear classifier along with a specificity and sensitivity of

85.6% and 76% respectively.

Gandhi et al. [15] developed an expert model for the detection of epileptic ac-

tivity in EEG signature using a probabilistic neural network for classification

of data. Clinical database was used for the training and testing the classifier.

Ten fold cross validation was used in the expert model. A high accuracy of

99.33% and specificity and sensitivity of 99.6% and 99% was obtained using

the model. Yuan et al. [16] compared the SVM and PNN classifier for de-

tection of seizure in EEG signal and showed that SVM classifier requires a

lower number of features sets compared to PNN to accurately determine the

data containing epilepsy. Cao’s method was used to compute the embedding

dimensions of the EEG data which were further used as feature sets for the

classifiers.


2.2 Previous research on prediction of seizures 6

2.2 Previous research on prediction of seizures

Prediction of seizure is a relatively new field as compared to the detection

which has been prevalent for a long time. In the late 1990s, a hypothesis [17]

[18] was given that the seizure starts building minutes to hours before the

actual attack. This led to an interest and subsequent work on the prediction

of seizure. However, till now, a robust technique has not been developed

to predict seizure in all situations. The papers on the prediction of seizures

are being reviewed by other authors as only near-chance predictions [19].We

came across many papers, such as those by Carney et al. [18] and Fazel-

Rezai et al. [20], where several techniques for seizure prediction have been

discussed. They provided a good starting point for the research.

Mormann et al. [21] showed that phase synchronization between different

EEG channels could be used as measure for detecting the pre-ictal stage. He

showed that the correlation coefficient can be used as statistical tool for the

finding the phase synchronization between the EEG channels. The reporting

shown in the paper shows that in 12 out of 14 cases, the ictal stages can be

predicted in advance. In another related paper, Mormann [22] found that

both the linear and non-linear techniques shows similar performance for the

prediction of seizures, thereby indicating that the changes in dynamics were

not necessarily caused solely by non-linear features of dynamics.

Le Van Quyen et al. [23] have used a method based on measuring the dynam-

ical similarity of the EEG recording with a reference state. It is a uni-variate

technique. The training data is first segmented into inter-ictal, pre-ictal and

ictal stages. For each stage, the method of delays has been used to trans-

form the time series data to a sequence of points in an m-dimensional space.

Instead of using the sample values directly, the duration between two con-

secutive positive zero crossings has been used. This allows for a greater

robustness to additive noise. Then, the dynamical similarity between any

two windows was calculated by computing the cross correlation integral be-

tween the windows. [24] The dynamical similarity is close to one if the signal

is stationary, while it is lower for non stationary signal. The signal windows

were compared to a reference window,extracted from the inter ictal stage.

The authors observed a decrease in similarity index, a few minutes before



the seizure occurred.

Maiwal et al. [25] have given a set of criteria and performance measures, to

evaluate seizure prediction techniques, before they can be applied clinically.

Their study mainly focuses in the most popular technique of dynamical sim-

ilarity [26].

Mormann et al. [19] states that many of the measures previously considered

suitable for prediction perform no better than a random predictor. He pro-

poses that only if the performance of a algorithm for the original seizure times

is significantly better than the performance for a number of independent re-

alizations of the surrogate seizure times, can the null hypothesis, namely,

that a given algorithm cannot detect a pre-seizure state with a performance

above chance level, be rejected. As an example, he specified that the paper

using correlation dimension, the Lyapunov exponent and the signal energy,

were not able to discriminate the pre-ictal from the inter-ictal period above

chance level.

Schad et al. [27] have investigated the detection and prediction of seizure, us-

ing scalp and intra-cranial recordings. They have used a neural network based

classifier, using simulated leaky integrate and fire neurons.Their method uses

the derivative of the EEG signal as a feature. The times at which this deriva-

tive exceeded a certain threshold is noted and a pulse is generated at this

time. These pulses are integrated with the leaky neurons, and when the in-

tegral value reaches a threshold, the integrator is reset to zero. A signal is

generated, every time the integrate and fire neuron is reset, giving a spiking

rate. This spiking rate is then used to declare the presence of seizure activ-

ity. In a similar way, the derivative of the spiking rate is employed to detect

pre-seizure changes, in order to predict the occurrence of seizure.

Mirowski et al. [28] have studied 4 bi-variate features of EEG synchroniza-

tion, namely, cross-correlation, interdependence, dynamical entrainment and

wavelet synchronization. Non linear interdependence measures the distance

between two EEG signals in state space, which is similar to dynamical sim-

ilarity. Dynamical entrainment is a measure of the difference between the

largest Lyapunov exponent between two time series, which is a measure of

chaos. The third is similar to phase synchronization technique, given in [22].

These have been used as features for several machine learning techniques,



like SVM, logistic regression and convolutional neural networks. They have

concluded that the convolutional neural network technique on the wavelet

coherence feature, achieves a sensitivity of 71%, without any false positives.

Zheng et al. [29] used a Bi variate Empirical Mode Decomposition(BEMD)

to first decompose windows of each pair of EEG to intrinsic mode func-

tions(IMF). BEMD was used due to the non-stationary nature of EEG sig-

nals. The BEMD technique decomposes the data into various IMFs without

losing the time information contained the original data. The mean phase

coherence for all the windows is then calculated and observed for occurrence

of pre-ictal stage. He observed that there was both an increase and decrease

of phase synchronization before the seizure onset.

Shufang et al. [30] used spike rate to distinguish the pre-ictal stage from ictal

stage. The research uses the hypothesis that with the advent of seizures, the

spike rate of various windows will take different values. They used a mor-

phology filtering operation to estimate the background EEG signal from the

original and deduct the background signal from the original signal to get the

spike signal. The morphology filter can be implemented with simple addition

and subtraction operations, thus reducing the computational complexity of

the algorithm. The spike rate of each window is analysed to find the increase

in the rate. Using the spike rate, they were able to predict a seizure with a

sensitivity of 75.8% and false prediction rate of 0.09hr. Guanghua et al. [31]

shows an improved spike detection algorithm in EEG signals using Morpho-

logical filtering operations.

Teixeira et al. [32] tried to classify the EEG signal of an epileptic patient

into 4 state: inter-ictal, pre-ictal, ictal and post-ictal. They used multiclass

SVM classifier and used 22 univariate features such as Energy, Mean, Power,

Hjorth Parameters, etc. They used an extensive patient database of 278 pa-

tient to perform this research. They were able to anticipate more than half

of the seizure through their method with a false prediction rate of less than

0.15hr. They also observed that invasive and non-invasive EEG recordings

gave nearly the same results in terms of prediction of seizures.


Chapter 3

Background Work

3.1 Electroencephalography

Electroencephalography (EEG) is the recording of electrical activity along

the scalp produced by the firing of neurons within the brain [33]. In clinical

contexts, EEG refers to the recording of the brain’s spontaneous electrical

activity over periods of time. EEG device is attached to the scalp through

electrodes. These electrodes are connected to a machine, which records the

electrical impulses. The signals are then processed to finally get the mental

state of the brain during the recording period. Each area of the brain pro-

duces a different brain wave according to the mental state, the kind of tasks

the person was imagining and the general condition of the patient [34].

The continuous monitoring of EEG signals can be highly helpful in clinical

environments e.g. for diagnosing epilepsy or brain tumor, predicting seizure,

detecting abnormal brain stages or even post stroke neuro-rehabilitation [35].

The EEG data is usually obtained through a EEG electrodes attached to

the scalp of the brain. These electrodes are connected to a machine, which

records the electrical impulses received from the brain.

The 10-20 system is an internationally for the placement of electrodes on

the scalp of the brain for recording the EEG data. The system is based on

the relation between the electrodes location and the corresponding area of

cerebral cortex. The number ’10’ and ’20’ are representative of the fact that

the distance between adjacent electrodes are either 10% or 20% of the total

front- back or left-right skull distance [4]. All the electrode placement sites

are given a specific name depending on the cerebral lobe underlying and the

hemisphere. The letters F, T, P and O stand for frontal, temporal, pari-

etal, and occipital lobes, respectively. The letter C is used to identify the

electrodes located over the central region of the brain. The even and odd

numbers used to refer to the electrodes points to the positioning of electrodes


3.2 EEG frequency bands 10

Figure 3.1: Brain Electrodes Location [4]

in right and left hemispheres respectively. Figure 1.1 shows the placement

of electrodes on the scalp for obtaining the EEG recording. The outer circle

in the figure is drawn at the level of nasion and inion while the inner circle

represents the temporal line of electrodes.

3.2 EEG frequency bands

The signals received through the BCI are then processed to finally get the

mental state of the brain during the recording period. Each area of the

brain produces a different brain wave according to the mental state, the kind

of tasks he was imagining and the general condition of the patient. The

EEG signal consists of different frequency bands defined by their wave to

characterize and identify psychological states [33]:

• Delta Waves: Frequency ranges up to 4 Hz. It tends to be the highest

in amplitude. It is seen normally in adults for slow wave sleep.

• Theta Waves: Frequency ranges from 4 Hz to 7 Hz. Theta is seen

normally in young children. It may be seen in drowsiness or arousal in

older children and adults; it can also be seen in meditation.


3.3 Wavelet Transform 11

• Alpha waves: Frequency ranges from 8 Hz to 12 Hz. It is higher in

amplitude on the dominant side.

• Beta Waves: Frequency ranges from 12 Hz to about 30 Hz. It is seen

usually on both sides in symmetrical distribution and is most evident

frontally. Beta activity is closely linked to motor behavior and is gen-

erally attenuated during active movements [15].

• Gamma Waves: Frequency ranges are approximately 30-100 Hz. Gamma

rhythms are thought to represent binding of different populations of

neurons together into a network for the purpose of carrying out a cog-

nitive or motor function.

The normal Electroencephalography (EEG) varies by age. The EEG in child-

hood generally has slower frequency oscillations than the adult EEG. Pre-

viously, Epileptic Seizures information were thought to be present only in

the lower frequency bands(Theta, Alpha, Beta). However, there are growing

number of studies which show that the high frequency oscillations(Gamma

rhythms) also contain important data about seizure activity [36]. In our

study, we have only used frequencies till 40 Hz for the detection and predic-

tion of epileptic seizures.

3.3 Wavelet Transform

Wavelet transform is a tool to study time varying signals. Unlike Fourier

transform, which represent the signal in frequency domain and loose the

time domain information, the wavelet transform is capable of representing

the frequency, as well as, time related information. In digital domain, where

signals are represented by their sampled values, the discrete wavelet trans-

form is used as an alternative to Fast Fourier transform [37]. We have used

a Haar Wavelet transform to get various frequency bands for the data. The

Haar wavelet transform can be computed in O(N) time, N being the size of

the input data. For a time series input of 2n values, the Haar wavelet trans-

form outputs two time series, each of length 2n−1[38]. One of these, called


3.4 Morphological Operations 12

Figure 3.2: One cycle of Haar wavelet transform

the approximate coefficients, are formed by taking the mean of two consec-

utive values in the original time series. The sum operation is analogous to

the low pass filtering of the data. The second series, called detail coefficients,

is formed by taking the difference of two consecutive values of the original

series. The difference operation as high pass filtering of the data. Since each

of the resultant time series have information bandwidth half of the original

time series, they are down-sampled by 2. One cycle transform is illustrated

in Fig 3.2.

3.4 Morphological Operations

Mathematical morphology is a tool for extracting image components useful

in the representation and description of region shape, such as boundaries,

skeletons and convex hulls[39]. Morphological operators often take a binary

image and a structuring element as input and combine them using a set

operator (intersection, union, inclusion, complement). They process objects

in the input image based on characteristics of its shape, which are encoded

in the structuring element. The structuring element is a small binary image,

i.e. a small matrix of pixels, each with a value of zero or one. The structuring

element is positioned at all possible locations in the image and it is compared

with the corresponding neighborhood of pixels. Some operations test whether



Figure 3.3: Erosion of an image [5]

the element fits within the neighborhood, while others test whether it hits

or intersects the neighborhood. The four basic operations of Morphological

filtering are[40]-[41]:

• Erosion : The erosion operation is defined as:

(fθg)(n) = minm=0,1,...,M−1

{f(n+m)− g(m)}

where n = 0, 1...., N −M . For an image, when the center of a struc-

turing element g(n) is located on the origin of the image f(n), then the

erosion of A by B is the locus of points reached by the center of B when

B moves inside A. This can be visualized in the Fig 3.3. The erosion

of the dark-blue square by a disk, results in formation of a light-blue

square.

• Dilation : The dilation operation is defined as:

(f ⊕ g)(n) = maxm=0,1,...,M−1

{f(n−m) + g(m)}

where n = M − 1,M − 2...., N − 1. For an image, when the center of

a structuring element g(n) is located on the origin of the image f(n),

then the dilation of A by B is the locus of the points covered by B when

the center of B moves inside A. This can be visualized in the Fig 3.4.



Figure 3.4: Dilation of an image [5]

The dilation of the dark-blue square by a disk results in formation of

the light-blue square with rounded corners.

• Opening : The opening operation is given by

(f ◦ g)(n) = [(fθg)⊕ g](n)

where f(n) and g(n) are discrete functions defined on F = 0, 1...., N − 1

and G = 0, 1....,M − 1 respectively and N >> M . For an image, the

opening operation is the locus of translations of the structuring element

B inside the image A. This can be visualized in the Fig 3.5. The opening

of the dark-blue square by a disk, results in formation of the light-blue

square with round corners.

• Closing : The closing operation is given by

(f • g)(n) = [(f ⊕ g)θg](n)

where f(n) and g(n) are discrete functions defined on F = 0, 1...., N − 1

and G = 0, 1....,M − 1 respectively and N >> M . For an image, the

closing is the complement of the locus of translations of the symmetric

of the structuring element outside the image A. This can be visualized



Figure 3.5: Opening of an image [5]

Figure 3.6: Closing of an image [5]

in the Fig 3.6. The closing of the dark-blue shape (union of two squares)

by a disk, which results in the union of the dark-blue shape and the

light-blue areas.


Chapter 4

Detection

In this chapter, we will describe our method for the detection of seizure in a

given EEG signal, the data sets we have used for testing the algorithm and

the results we have obtained for the same.

4.1 Bonn University Data

The dataset used for detection part was obtained from the University of Bonn

[12]. The dataset consists of 5 sets of data, labelled as A,B,C,D and E. Each

set consists of 100 EEG recordings, each recording for a duration of 23.6

seconds, at sampling rate of 173.61 Hz. All of the recordings were devoid of

artefacts, such as eye movements or muscle movement. The details of each

set are described below-

1. Set A and Set B correspond to 5 healthy patients’ surface EEG record-

ing. The set A consists of recording with eyes open, while set B consists

of recording with eyes closed.

2. Sets C and D consist of recording from patients suffering from epilepsy.

Set D corresponds to recordings from the epileptogenic zone, while

Set C consists of recordings from corresponding zone in the opposite

hemisphere. Both set C and set D were recorded during seizure free

intervals. That is, they correspond to inter-ictal stage.

3. Set E corresponds to data from patients with diagnosed epilepsy, recorded

during seizure intervals. That is, set E consists of the ictal stage.


4.2 Methodology 17

4.2 Methodology

In this section, we will explain the method used for the detection of seizure.

We have used Support Vector Machine Classifier to train and test the feature

sets extracted from each of the frequency bands. Discrete wavelet transform

was used. The procedure explained above has been implemented before in

Panda et al. [11] but only 3 features were used for each frequency band. The

following steps were implemented for the detecting the presence of seizure

activity in an EEG recording.

1. Pre-processing of data

2. Decomposition of Data into separate frequency bands

3. Feature set extraction

4. Training and testing of data

4.2.1 Pre-processing of Data

The data was pre-processed to remove high frequency (if any), noise in the

EEG as well as power line interference from the signal. The EEG recording

was filtered using an FIR filter with cutoff frequency of 40 Hz. A notch filter

at 50 Hz was applied to the EEG signal.

4.2.2 Decomposition of Data into separate frequency

bands

After pre-processing, a discrete time wavelet transform was used, to split

each EEG recording into the 5 sub bands, namely- delta, theta, alpha, beta

and gamma. The wavelet transform provides a convenient and simultaneous

picture of frequency and time for the signal. It also gives insights into the

significance of certain frequency bands that characterize a seizure. We have

used the Haar Wavelet transform. The Haar wavelet transform uses simple

mathematical operations for separating low and high frequency data. For low


4.2 Methodology 18

pass, it sums two consecutive values and for high pass, it takes difference of

consecutive values. The comparison of use of different wavelet transform was

done by Mohensi et al. [42]. Haar wavelet transform gave sufficiently good

results for decomposition of a signal into 2 bands and it was supplemented

by results too.

4.2.3 Feature Set Extraction

After separating the data into the five bands, the following features were

extracted from each of these bands:

1. The variance is the deviation of the signal from its mean. The variance

of seizure data is higher than the that of the normal data. It is given

as,

σ2 =

N∑i=1

(xi − µ)2

N

2. The mean signal energy of a seizure data generally tends to be higher

than that of normal data due to higher amplitudes. It is given as,

E =

N∑i=1

x2i

N

3. Power spectral density is calculated in two steps, first finding the

fast Fourier transform (X(wi)) of the time series and then taking the

squared modulus of the FFT coefficients.

P (wi) =|X(wi)|2

N

From this, the maximum and the minimum values are used.

4. Entropy is the measure of randomness and the information content of

a signal. To calculate the entropy of the given EEG signal, we have


4.2 Methodology 19

used the Shannon entropy formula-

Entropy = −N∑i=1

x2i log(x2

i )

This equation goves the un-normalized entropy of the signal.

5. The inter quartile range of the EEG signal gives the statistical dis-

persion of a signal, which is a measure of how squeezed or stretched

a distribution is. To calculate the inter quartile range, the signal is

divided into two parts, one containing values lower than the median

and the other containing values higher than the median. Then, the

interquartile range is the difference between the median of the left half

and median of the right half.

6. Kurtosis is a measure of how much weight the tail of the probability

distribution has. Thus, the presence of spikes is indicated by increased

kurtosis. Kurtosis is calculated as-

Kurt[X] =

N∑i=1

(xi − µ)4

Nσ4

where X denotes the time series of an EEG data set and N the number of

samples in the signal. X =(

x1 x2 x3 .... xN

). In all of these features,

described above, we aim to capture the presence of signal spikes, which is a

signature of an existing seizure activity.

4.2.4 Training the classifier

We then used a linear support vector machine, to classify each recording as

having seizure or not having seizure. Again, the linear SVM classifier has a

very simple model of calculations and is thus not very heavy computationally.

The SVM was initially trained on 240 samples, of which 80 were true positive

(containing seizure activity) and 160 were true negative (seizure free). Then,

the model was tested on 20 true positive data and 40 true negative data.


4.3 Testing & Results 20

Figure 4.1: Seizure free signal in the delta band

The source code for the above methodology can be found in the Appendix

section.

4.3 Testing & Results

Fig 4.1 shows the EEG signal for seizure free interval, while Fig 4.2 shows the

EEG signal for the ictal stage in the delta frequency band(0.5 to 4 Hz). As

we can see, the alpha band in the ictal stage contains high amplitude spikes,

as compared to that in the seizure free interval. The amplitude of the EEG

signal in the other bands also showed a rise in their values.

4.3.1 Feature set analysis

We have plotted the calculated values of the various features for the true

negative and true positive data. The seizure containing samples were found

to be significantly different from non-seizure samples for most of the the

frequency bands.



Figure 4.2: Seizure signal in the delta band

In the graphs Fig 4.3 - 4.6, the x-axis corresponds to the sample number,

while the y-axis corresponds to the entropy, maximum Power spectral density

and Kurtosis respectively. The x axis from 1 to 200 corresponds to sample

from seizure free intervals and from 201 to 300 correspond to samples having

presence of seizure activity.

1. Entropy: The entropy values in all bands declined, with Fig 4.3 de-

picting the entropy for wavelets in the sub band 10-20 Hz. We can see

that the entropy of the EEG samples in the ictal stage (samples from

201 onward) are lower than that of the seizure free stage (samples 1 to

200). Thus, we may conclude that during ictal stage, the randomness

in the EEG signal declines.

2. Power spectral density: We have plotted the maximum power spec-

tral density for the samples corresponding to the wavelets in the sub

band 10-20 Hz in Fig 4.4. Maximum PSD in the other sub-bands cor-

relate with this sub band. As expected, the maximum power spectral

density in during the ictal period (samples 201 to 300) is much higher



Figure 4.3: Entropy of samples for their wavelets in the alpha band

than that in the seizure free period (samples 1 to 200). Thus, power

spectral density is a good indicator of presence of seizure.

3. Kurtosis: The kurtosis for the samples is plotted in Fig 4.6 for wavelets

in the delta band and Fig 4.5 for those in the alpha band. As can be

seen, the kurtosis measure is relatively insignificant for classification in

the delta band. However, significant increase in kurtosis of the samples

is observed during ictal stage in the Alpha band wavelets. Thus, it

can be seen that some features are highly useful for classification in

some frequency bands while in other bands they are not particularly

important.

The signal energy, variance and inter-quartile range also showed significant

differences in the normal and the EEG sample. The signal energy and vari-

ance were higher for seizure duration due to high amplitude signals and the

inter-quartile range is higher due to presence of spikes in the signal.



Figure 4.4: Maximum Power spectral density for the samples correspondingto their wavelets in the alpha band

4.3.2 Accuracy, sensitivity, Specificity

Let us use TP for true positive classification, TN for true negative, FP for

false positive and FN for false negatives. The classification is TP if the

classifier detects a seizure, when the seizure is actually present, TN if it does

not detect seizure when seizure is absent, FP if it detects seizure when it

is actually absent and FN if it does not detect seizure when it is actually

present. The following evaluation measures were used-

1. Accuracy: It is defined as the sum of TP and TN as a percentage of

total test cases.

Accuracy =TP + TN

TP + TN + FP + FN∗ 100

2. Sensitivity: It is defined as -

Sensitivity =TP

TP + FN∗ 100


4.4 Discussion 24

Figure 4.5: Kurtosis of data samples corresponding to their wavelets in theAlpha band

3. Specificity It is defined as-

Specificity =TN

TN + FP∗ 100

Following our approach, we obtained an accuracy of 94.4%, sensitivity of

96.25% and specificity of 89.38%.

4.4 Discussion

Comparing with the existing methods, we can see that our algorithm for

seizure prediction performs well. It has a very high sensitivity and is thus

very good for detection of seizure activity in patients with diagnosed epilepsy.

The result is partly due to the large ensemble of features that we have used

for the classification purpose. However, another reason for significant im-

provement is the proper pre-processing of the data to remove artifact from

the EEG signal.


4.4 Discussion 25

Figure 4.6: Kurtosis of data samples corresponding to their wavelets in thedelta band

There could be further improvement in the results if more complicated tech-

niques were used for the wavelet decomposition of the signal. However, our

objective is to build a less computationally heavy algorithm that can be eas-

ily productized within low budget constraints. Therefore, the use of simple

techniques is important.


Chapter 5

Prediction

In this chapter, we have described the algorithm used for predicting a seizure

attack in advance using EEG signal. Previously we had showed a Proof

of Concept to demonstrate the potential of using the drop in Mean Phase

Coherence values as a reliable indicator of an impending seizure. We had a

small pool of test subjects on which we tested the algorithm.

Now, we have more than 10 test individuals data set on which we have tested

our algorithm. The algorithm is able to predict most of the epileptic seizures

more than 5 minutes before the actual onset. Through the literature review,

we found other feature which can be used to improve the Mean Prediction

Time of our algorithm. It was found that spike rate is be good indicator for

predicting seizures. It gave a reasonable result in most of the cases.

5.1 Data

The data sets used for prediction of seizure was obtained from 10 Epileptic

patients and 2 Normal Individuals. The EEG data was received for 18 brain

electrode channels, 9 in each hemisphere. The electrode placement was ac-

cording to International 10 20 standards. The recordings were sampled at

frequency of 200 Hz. 7 of the patients had cases of generalized epilepsy and

remaining 3 had a focal epilepsy. The data was obtained from the Grass

Telefactor Twin3 EEG machine at the Neurology and Sleep Center, Hauz

Khas, New Delhi.

The EEG data of an epileptic patient which contains seizures can be classified

into 4 major regions:

• Inter-ictal: Duration in-between the seizures and at sufficient gaps ( 3

hours) from any seizure activities.


5.2 Methodology 27

• Pre-ictal: Duration before and near to a seizure activities. Generally

chosen as 1-hour before a seizure. This is the region in which our system

should be able to predict that a seizure can occur in the near future.

• Ictal: The duration between the seizure beginning till its end.

• Post-ictal: Duration just after the seizure end till the EEG signal goes

back to normal.

5.2 Methodology

We tried 3 features to predict impending seizure activity. These were using

Mean - Phase Coherence, Spike rate and Bi-empirical mode decomposition.

Among these methods, the mean-phase coherence values between pairs of

neighboring electrodes gave the best results for the prediction. The Mean-

phase coherence values were found to decrease before the onset of an epileptic

seizure. The methodology and results for each of the algorithm are given

below. The source code for each of the methods can be found in the Appendix

section.

5.2.1 Mean Phase Coherence

Calculation of Mean Phase Coherence values between pairs of neighboring

electrodes has been used as a means of detecting phase synchronization be-

tween the 2 EEG channels. Phase Synchronization is defined as the locking

of the phases of 2 oscillatory system [51]. It is calculated as:

φn,m = nφa(t)−mφb(t) = constant (5.1)

The instantaneous phase of the signal can be calculated using

φ(t) = tan−1 s(t)

s(t)(5.2)


5.2 Methodology 28

where s(t) is the Hilbert transform of the signal given by

s(t) = −iFT−1(FT (S(t)) ∗ sign(w)) (5.3)

Here FT and FT−1 denotes the Fourier and the inverse Fourier transform

respectively of the signal.

To measure the relative phase between these two signals a and b and to

confine the phase between (−π/2, π/2), the following formula is used with

values of m = n = 1

φ1,1 = φa(t)− φb(t) = arctansa(t)sb(t)− sa(t)sb(t)sa(t)sb(t)− sa(t)sb(t)

(5.4)

The mean phase coherence is calculated as:

R =

∣∣∣∣∣N−1∑j=1

eiφ1,1(j∆T )

∣∣∣∣∣N

(5.5)

Studies show that the mean phase coherence [19] as a measure of phase

synchronization between the EEG channels changes significantly during the

pre-ictal stage when the brain waves shows signs of seizure formation as com-

pared to the inter-ictal stage [50]. This change in phase synchronization was

found to occur before the actual seizure of seizure. Therefore, by measuring

the value of mean phase coherence between channels, we were able to predict

the onset of seizures. The following steps were used:

• Pre-processing of data: The initial EEG data from 20 channels was

filtered with a band pass filter from 0.5 Hz to 40 Hz and a notch filter

at 50 Hz (to remove interference from the power supply of 50 Hz). The

data sampled at frequency of 200 Hz was divided into windows of 5

seconds each.

• Calculation of Mean phase coherence(MPC): The MPC values

were calculated for each of the windows between 2 pairs of electrodes.

The electrodes were chosen such only the coherence values between

the neighboring pairs of electrodes in each hemisphere were observed.


5.2 Methodology 29

This reduced the number of pairs to 24 from 153 (for 9 electrodes in

each hemisphere). The above step helps in reducing the computational

complexity as well as removes the redundancy in the data. The Hilbert

transform and MPC were then calculated using the equation 5.5.

• Thresholding and Classification By thresholding the running sum

of areas between the MPC values and their mean, we can find whether

a seizure attack is imminent in the near future. The threshold value

used in our case for 10 windows i.e. 50 seconds was -0.4. Whenever the

area between the graphs for the last 10 windows went below this value,

the alarm would be turned on to warn the patient of an impending

seizure. This threshold value was found to work fairly well for all the

cases.

5.2.2 Spike Rate

Seizures are usually characterized by generalized spikes in an EEG signal[31].

Spikes are transient signals with a pointed peak, clearly distinguishable from

the background activity. In addition, some previous studies [43],[44] have

shown that the spike rates change significantly minutes prior to a seizure.

Lange et al. [43] found that the rate of bilateral spikes increased obviously

across 20 min pre-ictal segments.

In this research, we have used spike rate as an indicator for prediction of

seizure. In an EEG signal, the spikes as well as the background EEG data

is present. [31] used Morphological filtering operations to identify the back-

ground noise and separate the spikes from the given EEG signal. We are also

going to use Mathematical Morphology to distinguish the spikes from the

original signal and thus, find the spike rate for the signal. The calculation is

as follows:

The opening and closing operations for Morphological filtering operations

were explained in 3.4. An opening operation can smooth the signal posi-

tive pulse (peak), while a closing operation can smooth the signal negative

pulse(valley). Since an ictal EEG signal contains both positive and neg-

ative phase spikes, the combination of morphological opening-closing(OC)


5.2 Methodology 30

and closing-opening(CO) are used to detect bi-directional spikes[31]. They

are defined as

OC(f(n)) = f(n) ◦ g(n) • g(n)

CO(f(n)) = f(n) • g(n) ◦ g(n)

Where f(n) is the input time series signal and g(n) is the structuring element.

However, since the filtering operations leads to a change in amplitudes, the

results of OC operations has lower amplitude than original signal and those

of CO operations have a higher amplitude. Thus, their weighted mean is

used to avoid amplitude detection.

y(n) = 0.5 ∗ (OC(x(n))) + CO(x(k)))

where

x(n) = y(n) + z(n)

Here, x(n) is the original windowed EEG signal. y(n) is the slow-changing

background signal and z(k) are the detected spikes. z(n) can be found by

subtracting y(n) from x(n).

z(n) = x(n)− y(n)

The structuring element g(n) used for the morphological operations deter-

mines the performance of the transform and thus the accuracy of the seizure

prediction. In order to remove the background components of original EEG

and highlight spikes, the structuring element selected should reflect the geo-

metrical characteristic of the original EEG signals, and at the same time, the

width of structuring element must be between spike rhythm and background

EEG rhythm. We select triangle structure element as follows:

g(n) = k ∗ (1− ‖n‖L

)

Here n varies from -L to L and hence, the total length of structuring element

is 2L+1. The duration of the spikes on the EEG varies from 20 to 70 ms,

and most of amplitudes exceed 100 [45]. In this project, the value of L has


5.3 Testing and Results 31

been chosen to be 5.

After detecting the spikes in the EEG signal, the spike rate is computed for

each 5 second window and results are compared to a threshold value which

can be used to classify the EEG signal window into pre-ictal, inter-ictal or

ictal stage.

5.2.3 Phase synchronization using bi-variate empirical

mode decomposition(BEMD)

In this method, the BEMD and Hilbert transform was used to decompose

pairs of EEG channels into intrinsic mode functions(IMFs) and extract phase

information from the data. The MPC values were then calculated for sepa-

rate windows for all the IMFs. The visual changes in each of the different

frequency bands were then observed and visualized for changes in the pre-

ictal stage.

5.3 Testing and Results

In this section, we will discuss the results obtained from using the three

techniques: Mean Phase Coherence, Spike Rate and BEMD. The first two

techniques, especially the first one, gave fairly good results for all the test

subjects. However, BEMD did not show good results for the data sets.

5.3.1 Mean Phase Coherence

In this part, We have presented our results for Seizure Prediction using Mean

Phase Coherence Values for 3 patients: one with focal epilepsy, with gener-

alized epilepsy and the last one for a normal patient. The value of mean

phase coherence was found to decrease significantly during the pre-ictal and

rise again during the ictal stages. This also proves the correctness of the

hypothesis that phase synchronization between different channels decreases

during pre-ictal stages. In the individual with no epilepsy, no such changes

were observed.



Figure 5.1: Seizure Activity for Patient 1

Patient with focal epilepsy

The first patient had a history of focal epilepsy in the occipital lobe. In

the recorded data available for the patient, the seizure is found at the time

00:13:01 i.e. 156th window for a window size of 5 seconds. Fig 5.1 shows the

location of seizure at the given location for all the channels.

Fig 5.2 shows the value of MPC between the channels O2 and T6 for all the

windows. The red line corresponds to the running mean values of the MPC

and the green line shows the smoothed values of the graph, smoothing done

by a moving average filter of length 4. Before the seizure actually occurs

at 156th window, we can see that the MPC values goes below the mean for

a long duration before finally rising back up at the seizure time. Though

there is variation of MPC throughout the recording period, the drop below

the mean just before the seizure is the longest and most widely pronounced.



Figure 5.2: MPC for patient 1 for channels O2 and T6

The MPC values just before seizure goes down around 93rd window. Around

window, we can sure due to continuous below mean values that this is the

pre-ictal stage. Hence, we could successfully predict seizures 4 minutes in

this case before the actual onset.

Fig 5.3 also shows the MPC values between O2 and P4. We can see that the

drop below the mean value is also pronounced highly just before the seizure.

However, in Fig 5.4, for the MPC values between T4 and F8, the drop is not

significant and thus, we will not able to tell from this data that a seizure is

impending. Therefore, we can say that the impact of phase synchronization

decrease in the pre-ictal period is not reflected in the entire brain in case of

a focal epilepsy.

Patient with generalized epilepsy

The second patient had a history of generalized epilepsy in the brain. The

data available is of 25 minutes. Fig 5.5 shows the seizure for the patient 2.

We can see that the EEG pattern shows large amplitude during the seizure

and also very small gaps between 2 seizure duration. The entire frame size is

around 100 seconds. The seizure is starting at 00:06:10. The pre-ictal stage



Figure 5.3: MPC for patient 1 for channels O2 and P4

Figure 5.4: MPC for patient 1 for channels F8 and T4



Figure 5.5: Seizure Window for patient 2

available for this patient is very short for only around 5 minutes. Hence, the

drop of the curve below the mean is not very distinctive. Fig 5.6 shows the

filtered MPC values between the channels C4 and P4. As we can see from

the figure, from the 10th window till the 55th window, there is a drop below

the mean of the filtered MPC values. Similarly, the above situation is also

observed for the MPC values(Fig 5.7) between channels C3 and P3 of the

left hemisphere.

Thus, we can receive an indication of the oncoming seizure from the MPC

graphs.

Normal Individual

We have also plotted the mean phase coherence for a normal person, while he

was asleep. The Fig 5.8 corresponds to the MPC between channel T4 and F8,



Figure 5.6: MPC for patient 2 for channels C4 and P4

Figure 5.7: MPC for patient 2 for channels C3 and P3



Figure 5.8: MPC for normal person for channels T4 and F8

Figure 5.9: MPC for normal person for channels F4 and FP2



while Fig 5.9 refers to the mean phase coherence between channel F4 and

FP2. From the graphs, we observe that the mean phase coherence values

neither increase, nor decrease considerably. Thus we can safely conclude the

absence of seizure. Also, there is no long term trend of MPC being below

the running average, so there is no prediction of an impending seizure. This

once again strengthens the basic idea of our approach.

5.3.2 Spike Rate

This subsection presents the results for utilizing the spike rate as a feature for

seizure prediction. The spike rate is seen to increase partially minutes before

the seizure. During the seizure activity, the spike rate reaches it maximum

values. It can also be seen the spike rate is different for different brain regions.

This concept can be used for focal localization of seizures.

Patient with seizure in specific regions

In this data set, the first seizure occurs around 12 minutes and 30 seconds

which corresponds to the 150th window in the EEG data. It is visible from

the seizure interval in Fig 5.10 that the seizure is present only in the left

hemisphere of the brain (Channels: PG1-P3). Fig 5.11 shows the Spike rate

of each window for the PG1 channel of Patient 1. It can be seen that though

the spike rate rises above 45 around 150th window, the increase in spike rate

begins from the 100th window, thus indicating an impending seizure. There is

a peak in spike rate in the end also which is more pronounced comparatively

in Fig 5.12 which represents spike rate in channel PG2 of right hemisphere.

The 1st seizure shows a very low spike rate of less than 10 in this channel

as compared to PG1. This is also an indicator of the fact that spike rate

is higher from those channels where the seizure is actually present. This

hypothesis can be used for the focal localization of a seizure and will be

highly helpful for the treatment.



Figure 5.10: Seizure Interval of Patient 1. Note that the seizure is present inleft hemisphere only



Figure 5.11: Spike Rate in Channel PG1(left hemisphere)

Figure 5.12: Spike Rate in Channel PG2(right hemisphere)



Figure 5.13: Spike Rate in Channel FP2

Patient with generalized epilepsy

The patient’s seizure in this case comes around 75th window as explained in

5.3.1 and shown in Fig 5.5. From the spike rate of Channel FP2 shown in

Fig 5.13, we can see that spike rate increases around 50th window and peaks

around 75th window. From the actual seizure Fig 5.5, we can see that not

all channels have similar peaks of amplitude during the seizures. For eg. the

channel PG1 has lower seizure amplitude peaks. The plot of this channel’s

spike rate Fig 5.14 also indicates a similar trend. The spike rate near 75th

window in PG1 is comparatively lower than those in channel FP2.

5.3.3 Bi-Empirical Mode Decomposition

As we can see from the Fig 5.15, the MPC values keep varying throughout the

duration of the data and there is no continuous decrease of the filtered value

of MPC below the running mean before the seizure duration (75th window).

Similar behavior is observed for other channels also. Since, the results we

achieved for this technique were not very distinctive, we did not use this

technique for the final results.



Figure 5.14: Spike Rate in Channel PG1

Figure 5.15: Plot of Mean phase coherence using BEMD for channels PG2and F8. The plot corresponds to the phase coherence values of 3rd IMF(intrinsic mode function)


5.4 Discussion 43

5.4 Discussion

The hypothesis that the phase synchronization decreased minutes before

seizure occurrence was found to be true for most of the cases. Mean Phase

Coherence was used as a measure of phase synchronization between EEG

channels. For testing, multi-channel EEG data was used for more than 10 test

subjects. During the pre-ictal stage, the MPC values were found to decrease

as compared to the inter-ictal stage. However, as the seizure approaches, the

MPC values rises sharply indicating a higher phase synchronization between

channels during ictal period and lower synchronization during the pre-ictal

period. The results were found to work in case of both focal and generalized

epileptic seizures. We also used spike rate as a feature for predicting the

onset of seizure. The spike rate was found to increase as the seizure activity

approaches. The prediction time in case of using spike rate as a measure is

lower than in the case of Mean Phase Coherence. Spike rate can also to be a

useful tool for source localization of an epileptic seizure. The last technique

which applied Bi Variate Empirical Mode Decomposition before computing

Mean Phase Coherence did not yield any fruitful results.


Chapter 6

Real Time Prediction Model

In this chapter, we describe the development of real-time model for the pre-

diction of epileptic seizure. The real-time model can predict an impending

seizure more than 5 minutes in advance with a fairly certain accuracy. It is

currently being implemented in MATLAB. However, it can be easily ported

to OCTAVE and then to Beagle Board/Raspberry Pi and be used for build-

ing a cost effective, point of care alert system. Such a device is not available

anywhere else in the world and it will be a revolutionary product in the field

of medical health care.

6.1 Methodology

In this model, we have used the concept of Mean Phase Coherence described

in subsection 5.2.1. However, now we are processing a single data window

at a time and analyzing it before proceeding to the next window. This way

we do not need the entire data at once. It is in contrast to the previous

implementation where the whole data was required in the beginning.

Three states were classified in the model: Inter-ictal, Pre-ictal and Ictal. The

states are colored with different colors in the results that we obtain: Blue for

Inter-ictal, Green for Pre-ictal and Red for Ictal. A threshold value was set

for area for fixed number of windows under the MPC curve and its running

mean curve. When the total area goes below this threshold in a particular

channel pair, the system gave an alert that a seizure attack might occur to

the patient in the next few minutes. The window size was kept at 2 sec for

all windows and the first few windows were neglected for calculating the area

to account for the sudden rise in MPC values.

The focus of the seizure may also be identified by carefully studying the

alarm occurrence time and MPC value drop in individual channel pair. We

found that for many cases that if seizure activity is visible in a particular


6.2 Results 45

region, then the MPC drop will be better visible in the channel pairs from

that region majorly. But there were anomalies with this concept for a few

cases as will be demonstrated below.

For the data, we have used the same data as used for the offline prediction

of seizure 5.1.

6.2 Results

In this section, we will show the results and inferences that we obtained from

using the above model. The results are shown for 2 patients.

6.2.1 Patient with Non-generalized Epilepsy

5.3.2 In this case, we observe from Fig 6.1 and 6.2 that the model was able

to detect the pre-ictal stage at 75th window. The actual seizure onset is on

150th window. This gives us a prediction time of more than 6 minutes. Also

as we saw in Figs 5.11 and 5.12, there are spikes in the end and hence, a

seizure activity is also happening in the end of the given data. This can also

be validated by the fact that the Mean Phase Coherence Values are dropping

below the threshold value.

6.3 Discussion

The real-time seizure prediction model in MATLAB as to classify the EEG

signal into 3 stages: inter-ictal, pre-ictal and ictal. The results we found

were pretty good. The area under the curve for a fixed number of windows

was thresholded to obtain this classification. The mean prediction time for

detecting the pre-ictal stage was around 6 minutes. The model was tested

on various test subjects and it predicted the seizure correctly for most of the

cases. A anomaly was observed in the case of the patient with generalized

epilepsy where 1 of the channel pairs did not show a decrease in the MPC

values whereas other channels were showing a decrease in phase synchroniza-

tion.


6.3 Discussion 46

Figure 6.1: Mean Phase Coherence values between channels T5 and P3. Bluedenotes inter-ictal stage, Green denotes pre-ictal, Red denotes ictal stage

Figure 6.2: Mean Phase Coherence values between channels T5 and O1


6.3 Discussion 47

In Fig 6.3, we can see 4 EEG channel pair getting processed parallely andshowing similar results. These channel pairs are from the left hemispherewhere seizure activity is actually occurring 5.10.

Figure 6.3: Mean Phase Coherence values for various channel pairs in lefthemisphere


6.3 Discussion 48

The Fig 6.4 show the MPC values and seizure stages as detected by the modelin the right hemisphere channels. It can be seen that there is no indicationof any ictal or pre-ictal activity in the graph near the 150th window. This iscorrect since the seizure activity is actually occurring in the left hemisphere.

Figure 6.4: Mean Phase Coherence values between channels FP2-F8 (righthemisphere)


6.3 Discussion 49

6.2.2 Patient with Generalized Epilepsy

5.3.1

Figure 6.5: Mean Phase Coherence values between channel pair related tooccipital zone

Figure 6.6: Mean Phase Coherence values between occipital zone and P4 innon-seizure hemisphere


6.3 Discussion 50

Figure 6.7: Mean Phase Coherence values between channel pairs(other thanoccipital lobe) in non-seizure hemisphere


Chapter 7

Conclusions and Future Work

In this project, we developed a real-time model for prediction of epileptic

seizures and implemented it in MATLAB.

The hypothesis that the phase synchronization decreased minutes before

seizure occurrence was found to be true for most of the cases. Mean Phase

Coherence was used as a measure of phase synchronization between EEG

channels. For testing, multi-channel EEG data was used for more than 10 test

subjects. During the pre-ictal stage, the MPC values were found to decrease

as compared to the inter-ictal stage. However, as the seizure approaches, the

MPC values rises sharply indicating a higher phase synchronization between

channels during ictal period and lower synchronization during the pre-ictal

period. The results were found to work in case of both focal and generalized

epileptic seizures. We also used spike rate as a feature for predicting the

onset of seizure. The spike rate was found to increase as the seizure activity

approaches. The prediction time in case of using spike rate as a measure is

lower than in the case of Mean Phase Coherence.

We also implemented a real-time seizure prediction model in MATLAB as

to classify the EEG signal into 3 stages: inter-ictal, pre-ictal and ictal. The

results we found were pretty good. The mean prediction time for detecting

the pre-ictal stage was around 6 minutes. The model was tested on various

test subjects and it predicted the seizure correctly for most of the cases. For

more accurate testing of the model, longer recordings with sufficiently large

inter-ictal stages shall be required so that there is no overlap of the effects of

two seizures occurring close to each other and leading to errors in the results.

The model developed by us for the prediction of seizures is highly promising

for conversion to an actual product in the market. The MATLAB code has

been ported to OCTAVE and the acquisition and processing of EEG signal

can now be easily performed on a Beagle Board/Raspberry Pi. Such a de-

vice for the prediction of seizure is not available anywhere else in the world

currently and it will be a revolutionary product in the field of medical health


52

care, benefiting millions of people without proper access to neuro clinical

facilities.


Bibliography

[1] U. R. Acharya, S. V. Sree, G. Swapna, R. J. Martis, and J. S. Suri, “Automated eeg

analysis of epilepsy: a review,” Knowledge-Based Systems, vol. 45, pp. 147–165, 2013.

[2] D. K. Binder and S. R. Haut, “Toward new paradigms of seizure detection,” Epilepsy

& Behavior, vol. 26, no. 3, pp. 247–252, 2013.

[3] B. Litt and J. Echauz, “Prediction of epileptic seizures,” The Lancet Neurology, vol. 1,

no. 1, pp. 22–30, 2002.

[4] G. H. Klem, H. O. Luders, H. Jasper, and C. Elger, “The ten-twenty electrode sys-

tem of the international federation,” Electroencephalogr Clin Neurophysiol, vol. 52,

no. suppl., p. 3, 1999.

[5] Wikipedia, “Mathematical morphology — wikipedia, the free encyclopedia,” 2016.

[Online; accessed 3-May-2016].

[6] W. H. Organization, The World Health Report 2001: Mental health: new understand-

ing, new hope. World Health Organization, 2001.

[7] R. A. Scott, S. D. Lhatoo, and J. W. Sander, “The treatment of epilepsy in developing

countries: where do we go from here?,” Bulletin of the World Health Organization,

vol. 79, no. 4, pp. 344–351, 2001.

[8] E. Magiorkinis, A. Diamantis, and K. Sidiropoulou, Hallmarks in the history of

epilepsy: from antiquity till the twentieth century. INTECH Open Access Publisher,

2011.

[9] J. Dorfmeister, M. Frei, D. Lerner, I. Osorio, and J. Ralston, “System for the predic-

tion, rapid detection, warning, prevention, or control of changes in activity states in

the brain of a subject,” Nov. 30 1999. US Patent 5,995,868.

[10] A. Aarabi and B. He, “A rule-based seizure prediction method for focal neocortical

epilepsy,” Clinical Neurophysiology, vol. 123, no. 6, pp. 1111–1122, 2012.

[11] R. Panda, P. Khobragade, P. Jambhule, S. Jengthe, P. Pal, and T. Gandhi, “Classifi-

cation of eeg signal using wavelet transform and support vector machine for epileptic

seizure diction,” in Systems in Medicine and Biology (ICSMB), 2010 International

Conference on, pp. 405–408, IEEE, 2010.

[12] R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, and C. E. Elger,

“Indications of nonlinear deterministic and finite-dimensional structures in time series

of brain electrical activity: Dependence on recording region and brain state,” Physical

Review E, vol. 64, no. 6, p. 061907, 2001.

[13] V. Srinivasan, C. Eswaran, and N. Sriraam, “Approximate entropy-based epileptic eeg

detection using artificial neural networks,” Information Technology in Biomedicine,

IEEE Transactions on, vol. 11, no. 3, pp. 288–295, 2007.


BIBLIOGRAPHY 54

[14] N. Ahammad, T. Fathima, and P. Joseph, “Detection of epileptic seizure event and

onset using eeg,” BioMed research international, vol. 2014, 2014.

[15] T. Gandhi, B. K. Panigrahi, M. Bhatia, and S. Anand, “Expert model for detection of

epileptic activity in eeg signature,” Expert Systems with Applications, vol. 37, no. 4,

pp. 3513–3520, 2010.

[16] Y. Yuan, “Detection of epileptic seizure based on eeg signals,” in Image and Signal

Processing (CISP), 2010 3rd International Congress on, vol. 9, pp. 4209–4211, IEEE,

2010.

[17] S. Ramgopal, S. Thome-Souza, M. Jackson, N. E. Kadish, I. S. Fernandez, J. Klehm,

W. Bosl, C. Reinsberger, S. Schachter, and T. Loddenkemper, “Seizure detection,

seizure prediction, and closed-loop warning systems in epilepsy,” Epilepsy & Behavior,

vol. 37, pp. 291–307, 2014.

[18] P. R. Carney, S. Myers, and J. D. Geyer, “Seizure prediction: methods,” Epilepsy &

Behavior, vol. 22, pp. S94–S101, 2011.

[19] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz, “Seizure prediction:

the long and winding road,” Brain, vol. 130, no. 2, pp. 314–333, 2007.

[20] A. Aarabi, R. Fazel-Rezai, and Y. Aghakhani, “Eeg seizure prediction: measures

and challenges,” in Engineering in Medicine and Biology Society, 2009. EMBC 2009.

Annual International Conference of the IEEE, pp. 1864–1867, IEEE, 2009.

[21] F. Mormann, K. Lehnertz, P. David, and C. E. Elger, “Mean phase coherence as a

measure for phase synchronization and its application to the eeg of epilepsy patients,”

Physica D: Nonlinear Phenomena, vol. 144, no. 3, pp. 358–369, 2000.

[22] F. Mormann, T. Kreuz, R. G. Andrzejak, P. David, K. Lehnertz, and C. E. Elger,

“Epileptic seizures are preceded by a decrease in synchronization,” Epilepsy research,

vol. 53, no. 3, pp. 173–185, 2003.

[23] M. Le Van Quyen, J. Martinerie, V. Navarro, P. Boon, M. D’Have, C. Adam, B. Re-

nault, F. Varela, and M. Baulac, “schad2008application,” The Lancet, vol. 357,

no. 9251, pp. 183–188, 2001.

[24] X. Li and G. Ouyang, “Nonlinear similarity analysis for epileptic seizures prediction,”

Nonlinear Analysis: Theory, Methods & Applications, vol. 64, no. 8, pp. 1666–1678,

2006.

[25] M. Winterhalder, T. Maiwald, H. Voss, R. Aschenbrenner-Scheibe, J. Timmer, and

A. Schulze-Bonhage, “The seizure prediction characteristic: a general framework to

assess and compare seizure prediction methods,” Epilepsy & Behavior, vol. 4, no. 3,

pp. 318–325, 2003.


BIBLIOGRAPHY 55

[26] M. Le Van Quyen, J. Martinerie, M. Baulac, and F. Varela, “Anticipating epileptic

seizures in real time by a non-linear analysis of similarity between eeg recordings,”

Neuroreport, vol. 10, no. 10, pp. 2149–2155, 1999.

[27] A. Schad, K. Schindler, B. Schelter, T. Maiwald, A. Brandt, J. Timmer, and

A. Schulze-Bonhage, “Application of a multivariate seizure detection and prediction

method to non-invasive and intracranial long-term eeg recordings,” Clinical neuro-

physiology, vol. 119, no. 1, pp. 197–211, 2008.

[28] P. Mirowski, D. Madhavan, Y. LeCun, and R. Kuzniecky, “Classification of patterns of

eeg synchronization for seizure prediction,” Clinical neurophysiology, vol. 120, no. 11,

pp. 1927–1940, 2009.

[29] Y. Zheng, G. Wang, K. Li, G. Bao, and J. Wang, “Epileptic seizure prediction using

phase synchronization based on bivariate empirical mode decomposition,” Clinical

Neurophysiology, vol. 125, no. 6, pp. 1104–1111, 2014.

[30] S. Li, W. Zhou, Q. Yuan, and Y. Liu, “Seizure prediction using spike rate of in-

tracranial eeg,” Neural Systems and Rehabilitation Engineering, IEEE Transactions

on, vol. 21, no. 6, pp. 880–886, 2013.

[31] G. Xu, J. Wang, Q. Zhang, S. Zhang, and J. Zhu, “A spike detection method in eeg

based on improved morphological filter,” Computers in biology and medicine, vol. 37,

no. 11, pp. 1647–1652, 2007.

[32] C. A. Teixeira, B. Direito, M. Bandarabadi, M. Le Van Quyen, M. Valderrama,

B. Schelter, A. Schulze-Bonhage, V. Navarro, F. Sales, and A. Dourado, “Epileptic

seizure predictors based on computational intelligence techniques: A comparative

study with 278 patients,” Computer methods and programs in biomedicine, vol. 114,

no. 3, pp. 324–336, 2014.

[33] E. Niedermeyer and F. L. da Silva, Electroencephalography: basic principles, clinical

applications, and related fields. Lippincott Williams & Wilkins, 2005.

[34] W. J. Freeman, “A neurobiological interpretation of semiotics: meaning, representa-

tion, and information,” Information Sciences, vol. 124, no. 1, pp. 93–102, 2000.

[35] D. Kumar, G. Aggarwal, R. Sehgal, A. Das, U. Lahiri, and A. Dutta, “Engagement-

sensitive interactive neuromuscular electrical therapy system for post-stroke balance

rehabilitation-a concept study,” 2015.

[36] S.-A. Lee, D. D. Spencer, and S. S. Spencer, “Intracranial eeg seizure-onset patterns

in neocortical epilepsy,” Epilepsia, vol. 41, no. 3, pp. 297–307, 2000.

[37] R. X. Gao and R. Yan, “From fourier transform to wavelet transform: A historical

perspective,” in Wavelets, pp. 17–32, Springer, 2011.

[38] D. H. Bailey and P. N. Swarztrauber, “The fractional Fourier transform and appli-

cations,” SIAM Rev., vol. 33, no. 3, pp. 389–404, 1991.


BIBLIOGRAPHY 56

[39] R. M. Haralick, S. R. Sternberg, and X. Zhuang, “Image analysis using mathematical

morphology,” Pattern Analysis and Machine Intelligence, IEEE Transactions on,

no. 4, pp. 532–550, 1987.

[40] P. Maragos and R. W. Schafer, “Morphological filters–part i: Their set-theoretic

analysis and relations to linear shift-invariant filters,” Acoustics, Speech and Signal

Processing, IEEE Transactions on, vol. 35, no. 8, pp. 1153–1169, 1987.

[41] P. Maragos and R. W. Schafer, “Morphological filters–part ii: Their relations to

median, order-statistic, and stack filters,” Acoustics, Speech and Signal Processing,

IEEE Transactions on, vol. 35, no. 8, pp. 1170–1184, 1987.

[42] H. R. Mohseni, A. Maghsoudi, and M. B. Shamsollahi, “Seizure detection in eeg

signals: A comparison of different approaches,” in Proc. of the 28th Annual Inter-

national Conference of the IEEE Engineering in Medicine and Biology Society, vol.

Supplement, 2006.

[43] H. H. Lange, J. P. Lieb, J. Engel, and P. H. Crandall, “Temporo-spatial patterns of

pre-ictal spike activity in human temporal lobe epilepsy,” Electroencephalography and

clinical neurophysiology, vol. 56, no. 6, pp. 543–555, 1983.

[44] W. Truccolo, J. A. Donoghue, L. R. Hochberg, E. N. Eskandar, J. R. Madsen, W. S.

Anderson, E. N. Brown, E. Halgren, and S. S. Cash, “Single-neuron dynamics in

human focal epilepsy,” Nature neuroscience, vol. 14, no. 5, pp. 635–641, 2011.

[45] A. A. Dingle, R. D. Jones, G. J. Carroll, and W. R. Fright, “A multistage system to

detect epileptiform activity in the eeg,” Biomedical Engineering, IEEE Transactions

on, vol. 40, no. 12, pp. 1260–1268, 1993.


Appendix A

Source code

A.1 Detection

1 f unc t i on [F ] = wavelet ( dr )

2 % This code imports the data in the f o l d e r s and

performs the wavelet %trans form . I t then c a l l s the

f e a t u r e e x t r a c t i o n code .

3 A = ze ro s (4097 ,100) ;

4 B = c e l l (5 ,100) ;

5 f o r i = 1:100

6 A( : , i ) = importdata ( dr ( i +2) . name) ;

7 [ h1 , l 1 ] = dwt (A( : , i ) , ’ haar ’ ) ;

8 [B{1 , i } , l 2 ] = dwt ( l1 , ’ haar ’ ) ;

9 [B{2 , i } , l 3 ] = dwt ( l2 , ’ haar ’ ) ;

10 [B{3 , i } , l 4 ] = dwt ( l3 , ’ haar ’ ) ;

11 [B{4 , i } , l 5 ] = dwt ( l4 , ’ haar ’ ) ;

12 [B{5 , i } , l 6 ] = dwt ( l5 , ’ haar ’ ) ;

13 end

14 F = ex t ra c t (B, 9 ) ;

15 end

1 f unc t i on [A] = e x t r a c t (B, n)

2 %This code e x t r a c t s a l l the f e a t u r e s mentioned in the

report , f o r the

3 % d e t e c t i o n part

4 A = ze ro s (n , 1 0 0 , 5 ) ;

5 f o r i = 1:100

6 f o r j = 1 :5

7 H = dsp . ZeroCross ingDetector ;


A.2 Prediction 58

8 A(1 , i , j ) = var (B{ j , i }) ;

9 A(2 , i , j ) = mean(B{ j , i }) ;

10 A(3 , i , j ) = wentropy (B{ j , i } , ’

shannon ’ ) ;

11 A(4 , i , j ) = i q r (B{ j , i }) ;

12 A(5 , i , j ) = k u r t o s i s (B{ j , i }) ;

13 A(6 , i , j ) = (A(1 , i , j ) + (A(2 , i , j

) ) ˆ2)∗ l ength (B{ j , i }) ;

14 PSD = periodogram (B{ j , i }) ;

15 A(7 , i , j ) = min (PSD) ;

16 A(8 , i , j ) = max(PSD) ;

17 A(9 , i , j ) = step (H,B{ j , i }) ;

18 end

19 end

20 end

A.2 Prediction

A.2.1 Mean Phase Coherence Calculation

1 f unc t i on [R] = p h a s e a l l ( data )

2 % This f i l e c a l l s the phasesyn func t i on to c a l c u l a t e

the MPC between

3 % the s e t o f p r ede f i n ed pa i r s , PS1 and PS2

4 PS1 = [ 1 , 1 , 1 , 5 , 5 , 8 , 8 , 4 , 6 , 3 , 3 ] ;

5 PS2 = [ 2 , 3 , 4 , 6 , 2 , 7 , 4 , 7 , 2 , 6 , 7 ] ;

6 Fs = 200 ;

7 w i n s i z e = 5 ;

8 no win = f l o o r ( s i z e ( data , 1 ) /( Fs∗w i n s i z e ) ) ;

9 R = ze ro s ( no win , s i z e (PS1 , 2 ) ) ;

10 f o r i = 1 : s i z e (PS1 , 2 )

11 % f o r j=i +1: s i z e ( data , 2 )

12 R( : , i ) = phasesyn ( data ( : , PS1( i ) ) , data ( : , PS2

( i ) ) , w i n s i z e ) ;


A.2 Prediction 59

13 end

14 end

1 f unc t i on [MPC] = phasesyn (T4 , F4 , w i n s i z e )

2 % THis func t i on implements the MPC c a l c u l a t i o n

a lgor i thm .

3 Fs = 200 ;

4 %w i n s i z e = 5 ;

5 no win = f l o o r ( s i z e (T4 , 1 ) /( Fs∗w i n s i z e ) ) ;

6 A = ze ro s ( no win , w i n s i z e ∗Fs ) ;

7 B = ze ro s ( no win , w i n s i z e ∗Fs ) ;

8 MPC = ze ro s ( ( no win ) ,1 ) ;

9 % d = f d e s i g n . lowpass ( ’N, Fc ’ , 10 , 100 , 200 ) ;

10 % Hd = des ign (d , ’ butter ’ ) ;

11 f o r i = 1 : ( no win )

12 A( i , : ) = T4( ( i −1)∗w i n s i z e ∗Fs+1 : i ∗w i n s i z e ∗Fs

) ;

13 B( i , : ) = F4 ( ( i −1)∗w i n s i z e ∗Fs+1 : i ∗w i n s i z e ∗Fs

) ;

14 A( i , : ) = A( i , : ) − mean(A( i , : ) ) ;

15 B( i , : ) = B( i , : ) − mean(B( i , : ) ) ;

16 % A( i , : ) = f i l t e r (Hd,A( i , : ) ) ;

17 % B( i , : ) = f i l t e r (Hd,B( i , : ) ) ;

18 % apply f i l t e r ( tape r ing )

19 hb A = imag ( h i l b e r t (A( i , : ) ) ) ;

20 hb B = imag ( h i l b e r t (B( i , : ) ) ) ;

21 ph = ze ro s ( w i n s i z e ∗Fs , 1 ) ;

22 f o r j = 1 : w i n s i z e ∗Fs

23 ph( j ) = atan ( ( hb A ( j )∗B( i , j ) − hb B ( j )∗A( i ,

j ) ) /(A( i , j )∗B( i , j ) + hb B ( j )∗hb A ( j ) ) ) ;

24 end

25 ph = ph ( ( w i n s i z e ∗Fs∗0.1+1) : w i n s i z e ∗Fs ∗0 . 9 ) ;

26 temp = exp (1 i ∗ph) ;

27 MPC( i ) = abs (mean( temp ) ) ;


A.2 Prediction 60

28 % hb A = hb A [ ( w i n s i z e ∗Fs∗0.1+1) : w i n s i z e ∗Fs

∗ 0 . 9 ] ;

29 end

30 % plo t (MPC) ;

31 end

A.2.2 Spike Rate Calculation

1 f unc t i on [ s p i k e r a t e ] = s p i k e r a t e (b)

2 k = 0 . 9 ;

3 s p i k e t h = 20 ;

4 Fs = 200 ;

5 w i n s i z e = 5 ;

6 no win = f l o o r ( s i z e (b , 2 ) /( Fs∗w i n s i z e ) ) ;

7 bkg = ze ro s ( s i z e (b , 1 ) , Fs∗w i n s i z e ) ;

8 s p i k e s = ze ro s ( s i z e (b , 1 ) , Fs∗w i n s i z e ) ;

9 data = ze ro s ( s i z e (b , 1 ) , Fs∗w i n s i z e ) ;

10 s p i k e r a t e = ze ro s ( s i z e (b , 1 ) , no win ) ;

11 f o r i = 10 : no win

12 data = b ( : , ( ( i −1)∗Fs∗w i n s i z e +1) : ( i ∗Fs∗w i n s i z e

) ) ;

13 bkg = bkg va l s ( data , k ) ;

14 s p i k e s = bkg−data ;

15 s p i k e r a t e ( : , i ) = d e t e c t s p i k e ( abs ( s p i k e s ) , Fs∗win s i z e , s p i k e t h ) ;

16 end

17 % plo t ( s p i k e r a t e ) ;

18 end

1 f unc t i on [ bkg ] = bkg va l s (b , k )

2 l s e =5;

3 bkg = ze ro s ( s i z e (b , 1 ) , s i z e (b , 2 ) ) ;

4 c l o s e = ze ro s (1 , s i z e (b , 2 ) ) ;

5 open = ze ro s (1 , s i z e (b , 2 ) ) ;

6 % se = ze ro s (2∗ l s e +1 ,1) ;


A.2 Prediction 61

7 spk th = 1 ;

8 NHOD = ones (1 ,2∗ l s e +1) ;

9 Height = 1−abs ( l i n s p a c e (−1 ,1 ,11) ) ;

10 se = s t r e l (NHOD, Height ) ;

11 f o r i = 1 : s i z e (b , 1 )

12 c l o s e = imc lo se (b( i , : ) , s e ) ;

13 open = imopen (b( i , : ) , s e ) ;

14 bkg ( i , : ) = 0 . 5∗ ( imopen ( c l o s e , se )+imc lo se ( open ,

se ) ) ;

15 end

16 end

1 f unc t i on [ s p i k e r a t e ] = d e t e c t s p i k e ( sp ikes , samples ,

s p i k e t h )

2 s p i k e r a t e = ze ro s ( s i z e ( sp ikes , 1 ) , 1 ) ;

3 f o r i = 1 : s i z e ( sp ikes , 1 )

4 f o r j = 2 : samples−1

5 i f ( ( s p i k e s ( i , j ) > s p i k e t h ) && ( s p i k e s ( i , j

)>s p i k e s ( i , j−1) ) && ( s p i k e s ( i , j )>s p i k e s (

i , j +1) ) )

6 s p i k e r a t e ( i ) = s p i k e r a t e ( i ) +1;

7 end

8 end

9 end

10 end

A.2.3 Bi-Empirical Mode Decomposition

1 f unc t i on [MPC]=calc MPC (IMF)

2 %Implements the MPC c a l c u l a t i o n

3 s e r i e s 1=r e a l (IMF) ;

4 s e r i e s 2=imag (IMF) ;

5 s e r i e s 1=s e r i e s 1−mean( s e r i e s 1 ) ;

6 s e r i e s 2=s e r i e s 2−mean( s e r i e s 2 ) ;

7 hb 1 = imag ( h i l b e r t ( s e r i e s 1 ) ) ;


A.2 Prediction 62

8 hb 2 = imag ( h i l b e r t ( s e r i e s 2 ) ) ;

9 ph = ze ro s ( s i z e ( s e r i e s 1 , 1 ) , 1 ) ;

10 f o r j = 1 : s i z e (ph , 1 )

11 ph( j ) = atan ( ( hb 1 ( j )∗ s e r i e s 2 ( j ) − hb 2 ( j )∗ s e r i e s 1 (

j ) ) /( s e r i e s 1 ( j )∗ s e r i e s 2 ( j ) + hb 1 ( j )∗hb 2 ( j ) ) ) ;

12 end

13 ph=ph( f l o o r ( s i z e (ph , 1 ) ∗0 . 1 ) : f l o o r ( s i z e (ph , 1 ) ∗0 . 9 ) ) ;

14 temp = exp ( (1 i )∗ph) ;

15 MPC = abs (mean( temp ) ) ;

1 f unc t i on [MPC] = bemd(T4)

2 % Implements the func t i on emd , which c a l c u l a t e s the

i n t r i n s i c mode func t i on

3 % of the data . Then c a l l s calc MPC , which c a l c u l a t e s

the MPC f o r g iven IMF %and pa i r o f channe l s .

4 Fs = 200 ;

5 w i n s i z e = 10 ;

6 N=4;%d four d i r e c t i o n s f o r two channe l s o f EEG

s i g n a l s

7 no win = f l o o r ( s i z e (T4 , 1 ) /( Fs∗w i n s i z e ) ) ;

8 num channel=s i z e (T4 , 2 ) ;

9 MPC=ze ro s ( no win , num channel , num channel , 2 0 ) ;

10 f o r i =1: no win

11 begin=1+(i −1)∗Fs∗w i n s i z e ;

12 l a s t=i ∗Fs∗w i n s i z e ;

13 %Find BEMD of two p a i r s

14 f o r j =1:3%num channel

15 f o r k=j +1:3%num channel

16 dat1=T4( begin : l a s t , j ) ;

17 dat2=T4( begin : l a s t , k ) ;

18 X=complex ( dat1 , dat2 ) ;

19 IMF = emd(X, ’NDIRS ’ ,4 ) ;

20 IMF=IMF ’ ;

21 f o r m=1: s i z e (IMF, 2 )


A.2 Prediction 63

22 MPC( i , j , k ,m)=calc MPC (IMF( : ,m) ) ;

23 end

24 end

25 end

26 end

27 end

A.2.4 Dynamical Similarity Index

1 f unc t i on [ cor ]=pred (A,B)

2 % This func t i on implements the dynamical s i m i l a r i t y

measure . I t i s %assumed that the i n i t i a l part i s

i n t e r i c t a l . We have j u s t implemented

3 % the measure o f s i m i l a r i t y o f two windows , as a

func t i on o f r , the %parameter f o r the gauss ian

d i s t ance between the t r a j e c t o r i e s .

4 %r e f e r e n c e=A

5 %t e s t=B

6 f s =200;

7 r e f w i n s i z e =20;

8 s l i d i n g w i n s i z e =10;

9 Nw=s l i d i n g w i n s i z e ∗ f s ;

10 Ns=r e f w i n s i z e ∗ f s ;

11 %Se l e c t ed channel− 3

12 dim = 9 ;

13 Ref=A( 1 : Ns) ;

14 c o u n t r e f = 0 ;

15 k=0;

16 Ref t=ze ro s (Ns , 1 ) ;

17 f o r i =1:Ns−1

18 i f ( ( Ref ( i ) <=0) & ( Ref ( i +1) >0) )

19 c o u n t r e f =c o u n t r e f +1;

20 Ref t ( c o u n t r e f ) = i−k ;

21 k = i ;


A.2 Prediction 64

22 end

23 end

24 r=std ( Re f t ) ;

25 R = l i n s p a c e ( r ∗0 .1 ,3∗ r , 1 00 ) ;

26 p h a s e r e f = ze ro s ( count r e f−dim+1,dim) ;

27 f o r i = 1 : dim

28 f o r p = 1 : ( count r e f−dim+1)

29 p h a s e r e f (p , i ) = Ref t (p+i −1) ;

30 end

31 end

32

33 [V, e i ] = e i g ( ( phase r e f ’ ) ∗( p h a s e r e f ) ) ;

34 X re f = p h a s e r e f ∗V;

35 [ a r e f , b r e f ]= s i z e ( X re f ) ;

36 num sl id win =20;%f l o o r ( s i z e (B, 1 ) / f s /Nw)

37 %s t a r t the windows

38 f o r j r e f =1: num sl id win

39 s l i d=B(1+( j r e f −1)∗Nw: j r e f ∗Nw) ;

40 s l i d t=ze ro s (Nw, 1 ) ;

41

42 j =0;

43 c o u n t s l i d = 0 ;

44

45 f o r i = 1 :Nw−1

46 i f ( ( s l i d ( i ) <=0) & ( s l i d ( i +1) >0) )

47 c o u n t s l i d = c o u n t s l i d + 1 ;

48 s l i d t ( c o u n t s l i d ) = i−j ;

49 j=i ;

50 end

51 end

52

53 p h a s e s l i d = ze ro s ( c o u n t s l i d−dim+1,dim) ;

54 f o r i = 1 : dim

55 f o r p = 1 : ( c o u n t s l i d−dim+1)


A.2 Prediction 65

56 p h a s e s l i d (p , i ) = s l i d t (p+i −1) ;

57 end

58 end

59

60

61

62 X st = p h a s e s l i d ∗V;

63 [ a , b ] = s i z e ( X st ) ;

64

65 Y re f = X re f ;%( 1 : ( min ( a r e f /2 , a ) ) , : ) ;

66 %ss imva l=ssim ( Y ref , X st ) ;

67 cor = ze ro s (100 ,1 ) ;

68 %cor2=ze ro s (100 ,1 ) ;

69 f o r r count =1:100

70 %c o r e l a t=co r r ( X st ’ , Y ref ’ ) ;

71 f o r i = 1 : a r e f%min ( a r e f /2 , a )

72 f o r j = 1 : a

73 c ro s s t e rm=exp(−norm ( ( Y re f ( i , : )−X st ( j , : ) ) ) /(R( r count ) ˆ2) ) ;

74 cor ( r count )=cor ( r count )+

cro s s t e rm ;

75 % cor2 ( r count )=cor2 ( r count )+(

c o r e l a t ( j , i ) /( a r e f ∗a ) ) ;

76 end

77 end

78 s e l f c o r r e f =0;

79 s e l f c o r s t =0;

80 f o r i =1: a r e f

81 f o r j =1: a r e f

82 t e rm re f=exp(−norm ( ( Y re f ( i , : )−Y re f ( j , : ) ) ) /(R( r count ) ˆ2) ) ;

83 s e l f c o r r e f=s e l f c o r r e f+te rm re f ;

84 end

85 end


A.3 Real Time Prediction Model 66

86 f o r i =1:a

87 f o r j =1:a

88 t e rm st=exp(−norm ( ( X st ( i , : )−X st ( j

, : ) ) ) /(R( r count ) ˆ2) ) ;

89 s e l f c o r s t=s e l f c o r s t+term st ;

90 end

91 end

92 cor ( r count )=cor ( r count ) / ( ( s e l f c o r s t ∗s e l f c o r r e f ) ˆ ( 0 . 5 ) ) ;

93

94 end

95 p lo t (R, cor ) ;

96 %hold a l l

97 %plo t (R, cor2 ) ;

98 %hold o f f

99

100 end

101 end

A.3 Real Time Prediction Model

A.3.1 Methodology

1 f unc t i on [MPC, data lp , data mean , data area ] = eeg pa r s e (

b , p lo t chan )

2 Fs=200;

3 w i n s i z e =5;

4 % mean windows = 50 ;

5 chan pa i r = 12 ;

6 no win = f l o o r ( s i z e (b , 2 ) /( Fs∗w i n s i z e ) ) ;

7 % mean windows = f l o o r ( no win /1) ;

8 data = ze ro s ( s i z e (b , 1 ) , Fs∗w i n s i z e ) ;

9 MPC = ze ro s ( chan pair , no win ) ;

10 f i l = 0.25∗ ones (1 , 4 ) ;



11 % data lp = ze ro s ( s i z e ( p lot chan , 2 ) , no win ) ;

12 % data mean = ze ro s ( s i z e ( p lot chan , 2 ) , no win ) ;

13 data lp = ze ro s ( chan pair , no win ) ;

14 data mean = ze ro s ( chan pair , no win ) ;

15 area win = 10 ;

16 data area = ze ro s ( chan pair , no win ) ;

17 % f o r i = 1 : s i z e ( p lot chan , 2 )

18 % hold o f f ;

19 % a x i s ( [ 0 2195 0 .7 1 ] ) ;

20 % hold on ;

21 c o l = 98∗ ones ( chan pair , no win ) ;

22 a r e a t h p r e = −0.4;

23 a r e a t h i c t a l = −0.00;

24 a r e a t h p o s t = 0 . 3 ;

25 % 98 f o r i n t e r i c t a l ( b lue ) , 103 f o r pre− i c t a l ( green ) ,

114 f o r i c t a l ( red )

26 pnt = 0 ;

27 i f ˜(˜ e x i s t ( ’ p lo t chan ’ , ’ var ’ ) | | isempty ( p lo t chan

) )

28 pnt = 1 ;

29 end

30 i f ( pnt == 1)

31 % f i g u r e

32 f o r j = 1 : s i z e ( p lot chan , 2 )

33 subplot ( s i z e ( p lot chan , 2 ) ,1 , j ) ;

34 a x i s ( [ 0 no win 0 .0 1 ] ) ;

35 hold on ;

36 % t i t l e ( s p r i n t f ( ’ Channel pa i r %g ’ , p lo t chan

( j ) ) ) ;

37 end

38 end

39 f o r i =1: no win

40 data = b ( : , ( ( i −1)∗Fs∗w i n s i z e +1) : ( i ∗Fs∗w i n s i z e

) ) ;



41 MPC( : , i ) = win proce s s ( data , Fs , w i n s i z e ) ;

42 f o r j = 1 : chan pa i r

43 data lp ( j , : ) = f i l t e r ( f i l , 1 ,MPC( j , : ) ) ;

44 % i f ( i <= mean windows )

45 data mean ( j , i ) = mean(MPC( j , 1 : i ) ) ;

46 % e l s e i f ( i > mean windows )

47 % data mean ( j , i ) = mean(MPC( j , ( ( i−mean windows ) : i ) ) ) ;

48 % end

49 i f ( i<=7+area win )

50 data area ( j , i ) = 0 ;

51 e l s e

52 data area ( j , i ) = sum( data lp ( j , ( i−area win ) : i ) )−sum( data mean ( j , ( i−area win ) : i ) ) ;

53 end

54 end

55 pause ( 0 . 0001 ) ;

56 i f ( ( i > 1) && ( pnt == 1) )

57 f o r j = 1 : s i z e ( p lot chan , 2 )

58 ch = plot chan ( j ) ;

59 subplot ( s i z e ( p lot chan , 2 ) ,1 , j ) ;

60 p lo t ( ( ( i −1) : i ) , da ta lp ( ch , i −1: i ) , ’ r ’ ) ;

61 p lo t ( ( ( i −1) : i ) , data mean ( ch , i −1: i ) , ’ b ’ )

;

62 i f ( data area ( ch , i ) < a r e a t h p r e )

63 c o l ( ch , i ) = 103 ;

64 e l s e i f ( ( c o l ( ch , i −1) == 98) && (

data area ( ch , i ) > a r e a t h p o s t ) )

65 c o l ( ch , i ) = 114 ;

66 e l s e i f ( c o l ( ch , i −1) == 98)

67 c o l ( ch , i ) = 98 ;

68 e l s e i f ( ( c o l ( ch , i −1) == 103) && (

data area ( ch , i ) > a r e a t h i c t a l ) )



69 c o l ( ch , i ) = 114 ;

70 e l s e i f ( ( c o l ( ch , i −1) == 103) && (

data area ( ch , i ) < a r e a t h i c t a l ) )

71 c o l ( ch , i ) = 103 ;

72 e l s e i f ( ( c o l ( ch , i −1) == 114) && (

data area ( ch , i ) > a r e a t h i c t a l ) )

73 c o l ( ch , i ) = 114 ;

74 e l s e i f ( ( c o l ( ch , i −1) == 114) && (

data area ( ch , i ) < a r e a t h i c t a l ) )

75 c o l ( ch , i ) = 98 ;

76 end

77 temp = [ data mean ( ch , i −1) data lp ( ch , i

−1) data lp ( ch , i ) data mean ( ch , i ) ] ;

78 temp2 = f i l l ( [ ( i −1) ( i −1) i i ] , temp ,

char ( c o l ( ch , i ) ) ) ;

79 s e t ( temp2 , ’ EdgeColor ’ , ’ none ’ ) ;

80 end

81 end

82 end

83 hold o f f ;

84 end

1 f unc t i on [R] = win proce s s ( data , Fs , w i n s i z e )

2 PS L1 = [11 , 11 , 16 , 12 , 16 , 13 , 13 , 13 , 18 , 18 , 18 , 14 ] ;

3 PS L2 = [16 , 12 , 12 , 13 , 17 , 14 , 18 , 17 , 17 , 14 , 15 , 15 ] ;

4 PS C1 = [ 1 1 , 1 6 , 1 7 , 1 8 , 1 5 ] ;

5 PS C2 = [ 2 , 7 , 8 , 9 , 6 ] ;

6 PS R1 = [ 2 , 2 , 7 , 3 , 7 , 4 , 4 , 4 , 9 , 9 , 9 , 5 ] ;

7 PS R2 = [ 7 , 3 , 3 , 4 , 8 , 5 , 9 , 8 , 8 , 5 , 6 , 6 ] ;

8 channe l s = s i z e (PS R1 , 2 ) ;

9 R = ze ro s ( channels , 1 ) ;

10 f o r i = 1 : channe l s

11 R( i ) = win MPC( data ( PS L1 ( i ) , : ) , data ( PS L2 ( i )

, : ) , Fs , w i n s i z e ) ;



12 end

13 end

1 f unc t i on va l = win MPC(D1 , D2 , Fs , w i n s i z e )

2 %D1, D2 are the windows o f 2 channe l s

3 D1 = D1−mean(D1) ;

4 D2 = D2−mean(D2) ;

5 hb 1 = imag ( h i l b e r t (D1) ) ;

6 hb 2 = imag ( h i l b e r t (D2) ) ;

7 ph = ze ro s ( w i n s i z e ∗Fs , 1 ) ;

8 f o r j = 1 : round ( w i n s i z e ∗Fs )

9 ph( j ) = atan ( ( hb 1 ( j )∗D2( j ) − hb 2 ( j )∗D1( j ) ) /(

D1( j )∗D2( j ) + hb 1 ( j )∗hb 2 ( j ) ) ) ;

10 end

11 ph = ph ( ( w i n s i z e ∗Fs∗0.1+1) : w i n s i z e ∗Fs ∗0 . 9 ) ;

12 temp = exp (1 i ∗ph) ;

13 va l = abs (mean( temp ) ) ;

14 end

1 f unc t i on plot sum (MPC, data lp , data mean , p lot chan ,

no win )

2 hold o f f ;

3 sum lp=ze ro s (1 , no win ) ;

4 sum mean=ze ro s (1 , no win ) ;

5 f o r i = 1 : s i z e ( p lot chan , 2 )

6 sum lp = sum lp + 10∗ data lp ( p lo t chan ( i ) , : ) ;

7 sum mean = sum mean + 10∗data mean ( p lo t chan ( i )

, : ) ;

8 end

9 p lo t ( sum lp ) ;

10 hold a l l ;

11 p lo t ( sum mean ) ;

12 end


Real-time Prediction of Epileptic Seizures using EEG...

Documents

Transcript of Real-time Prediction of Epileptic Seizures using EEG...