FPGA IMPLEMENTATION OF HUANG HILBERT TRANSFORM FOR CLASSIFICATION OF EPILEPTIC SEIZURES USING ARTIFICIAL NEURAL NETWORK

International Journal of VLSI design & Communication Systems (VLSICS) Vol.10, No.3, June 2019
DOI : 10.5121/vlsic.2019.10302 19
FPGA IMPLEMENTATION OF HUANG HILBERT
TRANSFORM FOR CLASSIFICATION OF EPILEPTIC
SEIZURES USING ARTIFICIAL NEURAL NETWORK
G.Deepika1
and K.S.Rao2
1
Research scholar at JNTU,Hyderabad , Asso.Prof at RRS college of Engg,
2
Director & Professor in ECE Dept,Anurag group of Institutions, Hyderabad
ABSTRACT
The most common brain disorders due to abnormal burst of electrical discharges are termed as Epileptic
seizures. This work proposes an efficient approach to extract the features of epileptic seizures by
decomposing EEG into band limited signals termed as IMF’s by empirical decomposition EMD. Huang
Hilbert Transform is applied on these IMF’s for calculating Instantaneous frequencies and are classified
using artificial neural network trained by Back propagation algorithm. The results indicate an accuracy of
97.87%. The algorithm is implemented using Verilog HDL on Zynq 7000 family FPGA evaluation board
using Xilinx vivado 2015.2 version.
KEYWORDS
EEG, IMF,EMD
1. INTRODUCTION
About 1 to 2% of world population is suffering from Epilepsy which is a neurological disorder.
The excessive synchronization of cortical neuronal networks results in sudden recurrent and
transient disturbances of perception or behavior. This neurological condition is chronic abnormal
bursts of electrical discharges in brain experienced by an individual and these recurrent seizures
are termed as Epileptic seizures.
For clinical manifestation these seizures are divided into partial, focal, generalized, unilateral and
unclassified. Electrical disturbances that occur only in part of central hemisphere and produce
symptoms related to mental functions are named as Focal epileptic seizures. Symptoms of
bilateral motor functions with loss of consciousness involving entire brain are termed as
generalized seizures. People at all ages can experience these two. EEG plays an important role in
diagnosing Epilepsy. Other unwarranted signals like power line interference and noise termed as
artifacts often distort EEG. Several techniques are available to identify and separate or suppress
these artifacts from EEG and the most common is Independent Component Analysis (ICA). The
performance of ICA technique is not well for muscle artefacts which are highly non-stationary.
For detecting and classifying epileptic seizures various time-frequency methods have been
proposed most importantly the wavelet transform and all these are nonstationary methods [4].
Various non-linear methods like probabilistic methods such as entropies, fractal dimensions,
lyapunov exponents are also being utilized in parallel [5]. For analyzing non stationary data a new
method was proposed in 1998 by N.E.Huang et al known as Huang Hilbert transform HHT [3][6].
Empirical Mode Decomposition technique converts any random signal into sum of Intrinsic Mode
Functions called IMF’s and a residue [8]. For noise removal and detrending purposes EMD can

20
be utilized based on the characteristics of IMF’s and those IMF’s corrupted by noise should not
be considered [9][10].
This paper presents a method to extract features that are being selected from IMF’s based on
weighted mean Frequency of each IMF and are classified using Neural Networks. The algorithm
is implemented on FPGA.
2. METHODOLOGY
University of Bonn EEG database is used in this work that is available in public domain in three
different cases
1) Healthy
2) Subjects during the interval free from seizures (inter-ictal) and
3) Subjects during the seizure (ictal) [7].
Single channel EEG data was taken from BONN database consisting of 100 recordings of 23.6
sec duration segments. Each segment is sampled at a rate of 173.61Hz and are digitized by 12 bit
ADC consisting of 4096 samples.
A. Hilbert-Huang Transform (HHT)
This technique is critically important in analyzing nonlinear and nonstationary data and more
importantly for representing time frequency energies. Such a system can better be described in
terms of instantaneous frequency which in turn reveals the intra wave frequency modulations and
Hilbert transform is the easiest way to measure the instantaneous frequency. Hilbert transform
can be used to define analytic signal in time series containing imaginary part as amplitude and
phase as time dependent. Any oscillations whether Linear or Non Linear that has equal number of
extreme values and zero crossings and symmetric to local mean are termed as intrinsic mode
oscillations and are represented by time series.This resulted in the usage of Hilbert transform to
only narrow band passed signals with the equal number of extreme values and zero crossings. To
avoid the alterations in harmonics which in turn creates distortion in the signal by filtering
operation which is linear, an Empirical Mode Decomposition (EMD) introduced by N.Huang in
1998 [3] is used before applying Hilbert transform. A comparison of various transformations is
given in Table -1.
Table-1: Comparison of Transformations.
FOURIER WAVELET HHT
Basis A priori A priori adaptive
Frequency Convolution, Global
uncertainity
Convolution, Global
uncertainity
Differential, Local
certainity
Presentation Energy-Frequency Energy Time
Frequency
Energy Time
Frequency
Non Linearity No No Yes
Non stationary No Yes Yes
Feature Extraction No Discrete: No
Continuous : Yes
Yes
The transform is adaptive and was defined from decomposition. The intrinsic mode functions are
defined by satisfying two conditions:
1) The difference between the number of extrema’s and the zero crossings should not be more
than one and

21
2) The average of maximum and minimum envelopes should be zero as defined by local
average.
Time series signal is represented by intrinsic modes of oscillations and each such oscillation
whether linear or nonlinear represents a simple oscillation that has equal number of extreme
values and zero crossings and also symmetric to local mean. These are called intrinsic mode
functions (IMF’s) that are mono components. The methodology to acquire IMF is called sifting
and is explained below:
1. All the Maxima and Minima of x(t) should be identified.
2. Use cubic spline interpolation to produce upper Xup(t) and Xlow(t) envelopes
3. Calculate the mean, point by point of envelopes with
m(t) = Xup(t) + Xlow(t) /2
4. Calculate the deviation d(t) = x(t)-m(t)
5. If d(t) equals the conditions of IMF as defined previously the IMF is achieved and then
replace x(t) with the difference residue r(t)= x(t)- d(t)
6. Otherwise there is no IMF and replace x(t) with d(t).
7. All the above cited five steps are repeated following the stopping
8. criteria till a single Maximum or Minimum residue signal is achieved.
= ∑
| |
(1)
The signal x(t) at the end is expressed as
= ∑ + !
"
# (2)
where rN(t) is the final residue and can be treated as DC component and Cj(t) are the INF’s with
zero means and are orthogonal to each other.
Hilbert Transform:
A real signal is represented in complex form as
$ = + %& (3)
Where XH(t) is Hilbert Transform of x(t) defined as
' +
#
(
)* +
,
,
-
∞
∞ (4)
the cauchy principal value of the integral is represented by P. The exponential form of Eq. (2) is
written as
$ = . . 01 (5)
2ℎ . = 4% 5 + %'
5 (6)
1 =
67
6
(7)
The time derivative of (5) is

22
$ = . . 8
. 2 + 8
. .1 (8)
Where w(t) is the time derivative of angular frequency known as instantaneous angular
frequency.
W t = ψ t =
<
<=
arctanB
CD =
C =
E (9)
The instantaneous frequency is defined as
F =
G
5(
(10)
The weighted mean frequency(MF) is computed from instantaneous frequencies
HI =
∑ J 0 .K 0L
MN
∑ J 0 .K 0L
MN
(11)
A1 & f1 – instantaneous amplitude and frequency, and the following are the features extracted
from each IMF
1) Rate of change of amplitude envelopes of IMF
2) Weighted mean frequency
3) Inter quartile range in each IMF defined as
IQR = Q3 – Q1 (12)
where Q1 and Q3 are the first and third quartiles.
For a set of data the mean of absolute deviations about the data’s mean is defined as Mean
absolute Deviation (MAD) given by
#
!
∑ | 0 − |!
0 # (13)
3. RESULTS AND DISCUSSION
The instantaneous frequency of the corresponding IMFs and the extracted IMFs of seizure signal
and non- seizure signal data are shown in Fig 1 and 2 respectively.

23
Fig.:1 seizure signal.
Fig.:2 Non-seizure signal.
The mean instantaneous frequencies of seizure and non-seizure data IMFs are given in Table -2.
Table 2: IMF’s Data of Seizure and Non Seizure signal.
Intrinsic Models Weighted Freq. Non-seizure Data Weighted Freq. Seizure Data
1 12.63 5.65
2 5.62 2.45
3 0.64 0.54
4 0.26 0.28
5 0.20 0.17
6 0.05 0.05
The following features are extracted from the selected IMFs from epoch of 5 seconds:
1) Mean weighted frequency
2) Rate of change of envelope amplitude
3) Inter Quartile Range
4) Mean absolute deviation.
The IMFs weighted mean frequency discriminates the seizure and non- seizure data and from the
above table 2 to 4 IMFs are useful for extracting the above features. IMF’s from 2 to 4 are being
considered and 12 features are extracted from 400 epochs of seizure and non-seizure data. A
neural network with 12 input neurons, one hidden layer with 15 neurons and one output neuron is
designed and a feed forward back propagation algorithm is implemented. The ANN model is as
shown below in Fig 3:

24
Fig 3: Neural Network Architecture.
The network is trained and simulated with 400 epochs of seizure and non- seizure data and a
result of one false negative and zero false positives in one epoch is achieved. The results are
tabulated in table 3 mentioned below:
Table 3: Seizure and Non Seizure Classification Results.
TP TN FP FN SP SE TCA
399 400 0 01 100% 99.75% 99.8%
Where
TP is True Positive, TN is True Negative
FP is False Positive, FN is False Negative
SE sensitivity = Number of TP decisions / No.of Positive cases
SP Specificity = No.of TN decisions /No.of Negative cases
TCA Total Classification accuracy = No.of Correct decisions /Total
no.of cases.
From the above results it is evident that the SP,SE and TCA of the proposed method is good and
superior.The algorithm is developed using Verilog HDL using Xilinx Vivado and implemented
Zync 7000 series family FPGA evaluation board. The synthesized and elaborated RTL schematics
are shown in Fig. 4 and 5 respectively.

25
Fig 4: Synthesized schematic.
Fig 5: RTL schematic.
The summary report of synthesis shows the less number of usage of LUTs and cells as given in
Fig.6.

26
Fig. 6: Synthesis report.
4. CONCLUSIONS
This work analyses the seizure and non-seizure EEG signals and IMFs are extracted by empirical
mode decomposition of the signals. IMFs based upon weighted mean frequency are selected for
feature extraction and are classified using neural network model with feed forward back
propagation algorithm. The proposed technique provides a tool for neuro physicians for
identifying brain abnormalities due to epilepsy and also can also be developed as ASIC. The
proposed technique essentially provides benefits like fast diagnosis, high accuracy with good
sensitivity and specificity.
ACKNOWLEDGEMENTS
The authors are highly thankful to the Chairman of ANURAG GROUP OF INSTITUTIONS,
Dr.P.Rajeshwar Reddy for his constant encouragement and also providing all the necessary
resources to carry out this work. They also thank Mr.Ramakrishna, Asst.Prof. Dept. of ECE,
Anurag group of Institutions for his suggestions.
REFERENCES
[1] J. Gotman., “Automatic recognition of epileptic seizures in the EEG,” Clinical Neurophysiology, vol.
54, pp. 530–540, 1982
[2] J.Gotman., “Automatic seizure detection: improvements and evaluation,” Clinical Neurophysiology,
vol. 76, pp. 317–324, 1990
[3] N.E. Huang, Z. Shen, S.R. Long, M.L. Wu, H.H. Shih,Q. Zheng, N.C. Yen, C.C. Tung, and H.H. Liu,
“TheEmpirical Mode Decomposition and Hilbert Spectrumfor Nonlinear and Nonstationary Time
Series Analysis,” Proc. Roy. Soc., vol. 454, pp. 903 – 995, 1998.
[4] Y.U. Khan, J. Gotman, “Electroencephalogram Wavelet based automatic seizure detection intra
cerebral”, Clinical Neurophysiology, vol. 114, pp. 899-908, 2003

27
[5] Güler NF, Übeyli ED, Güler.”Recurrent neural networks employing Lyapunov xponents for EEG
signal classification”,Expert Syst Appl. 2005; 29(3):506-14
[6] Varun Bajaj, Ram Bilas Pachori “Epileptic Seizure Detection Based on the Instantaneous Area of
Analytic Intrinsic Mode Functions of EEG Signals,” Biomed Eng Lett, vol. 3, pp. 17-21, 2013
[7] EEG time time series (epilepticdata)(2005,Nov.) [Online],
http://www.meb.unibonn.de/epileptologie/science/physik/eegdata.html
[8] Hedi Khammari , Ashraf Anwar, “A Spectral Based Forecasting Tool of Epileptic Seizures ” IJCSI
International.Journal of Computer Science Issues, Vol. 9, Issue 3, No 3, May 2012
[9] Rami J Oweis and Enas W Abdulhay., “Seizure classification in EEG signals utilizing Hilbert- Huang
transform” BioMedical Engineering OnLine 2011, 10:38
[10] lajos losonczi, lászló bakó, sándor-tihamér ,Brassai and lászló-ferenc Márton., “Hilbert-huang
transform used for eeg signal analysis ,” The 6th edition of the Interdisciplinarity in Engineering
International Conference , “Petru Maior” University of Tîrgu Mure, Romania, 2012
AUTHORS PROFILE
G. Deepika obtained her B.E., M.Tech degree in ECE in the year of 2002, 2005 from CBIT,
Osmania University and JNT University, Hyderabad respectively. She had 10 years of
teaching experience. Presently she is an Associate Professor in RRS College of Engineering
& Technology, Medak District and pursuing Ph.D in JNT University, Hyderabad.
Dr. K. S. Rao obtained his B. Tech, M. Tech and Ph.D. in Electronics and Instrumentation
Engineering in the years 1986, 89 and 97 from KITS, REC Warangal and VRCE Nagpur
respectively. He had 25 years of teaching and research experience and worked in all
academic positions, presently he is the Director, Anu rag Group of Institutions
(Autonomous) Hyderabad. His fields of interests are Signal Processing, Neural Networks
and VLSI system design.

FPGA IMPLEMENTATION OF HUANG HILBERT TRANSFORM FOR CLASSIFICATION OF EPILEPTIC SEIZURES USING ARTIFICIAL NEURAL NETWORK

More Related Content

Similar to FPGA IMPLEMENTATION OF HUANG HILBERT TRANSFORM FOR CLASSIFICATION OF EPILEPTIC SEIZURES USING ARTIFICIAL NEURAL NETWORK (20)

Recently uploaded (20)

FPGA IMPLEMENTATION OF HUANG HILBERT TRANSFORM FOR CLASSIFICATION OF EPILEPTIC SEIZURES USING ARTIFICIAL NEURAL NETWORK