Detection of Epileptic Seizures using Statistical Features in the EMD Domain S M Shafiul Alam Student No: 0409062207 Department of Electrical and Electronic Engineering Bangladesh University of Engineering and Technology
Detection of Epileptic Seizures using Statistical Features in the EMD Domain
S M Shafiul AlamStudent No: 0409062207
Department of Electrical and Electronic EngineeringBangladesh University of Engineering and Technology
Outline Background Motivation Objective Methodology Results and Discussion Conclusion Future Research Questions and Answer
2
BACKGROUND
3
Human Brain
4
Complex network of approx. 100 billion neurons
Information processing Communication with and within rest of the
body
Human Brain
5
Neuron – fundamental element of the network Interconnected using Axon and Dendrite Transmission of electric impulse
Synapse
Epilepsy and Seizure
Most common and serious neural disease Results from brain malfunction Causes recurrent seizure activity
6
Epilepsy and Seizure
Violent body shacking Muscle twitching Lost of consciousness
7
Unpredictable Unexpected injury Death
EEG Electroencephalogram Records neural activity from electric impulses Non-stationary Non-linear
8
Scalpelectrode
Subduralelectrode
Depthelectrode
9
Normal Pre-Seizure
SeizurePost-Seizure
Detection of Seizure
10
Continuous Monitoring
How long is the EEG records? How many patients are there? Hospitalization for how long?
Automatic System
Alternative to continuous monitoring
Gives alarm when seizure occurs Enables automated drug delivery Less hospitalization
11
MOTIVATION
12
Automatic EEG Monitoring System
Diagnosis Is the person suffering from epilepsy?
13
Detection Is the person experiencing a seizure?
General Strategy
14
EEG records
Pre-processing
Feature extraction
Feature based classification
Pre-processingMandatory Noise suppression Artifact reduction (e.g. muscle activity, heart activity, eye activity)
15
OptionalTransform into another domain (e.g. FT, DWT, TFA etc.) Power spectrum estimation
Feature extraction
Energy content Statistical feature (e.g., mean, standard deviation, variance etc.) Chaotic feature (e.g. LE, CD, FD, AE etc.)
16
Feature based classification
17
Threshold based decisionLinear classifierNaïve BayesKNNANNSVM
Issues
Classification performance Computational requirement Speed Practical implementation
18
OBJECTIVE
19
Transform EEGs into EMD domain Extract EMD domain statistical and chaotic
features Investigate features’ discriminating
capability Design ANN based classification system Study performance for epilepsy diagnosis
and seizure detection
20
EMD Empirical Mode Decomposition Output – set of IMFs Independent of linearity/non-linearity Independent of stationarity Adaptive Requires no basis function (Different from FT, DWT and other
TFA)21
Sifting Process
22
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
Local max. ()
Sifting Process
23
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
Envelope (max.)
Local min. ()
Envelope (max.)
Sifting Process
24
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
Envelope (max.)
Envelope (min.)
Mean
Sifting Process
25
10 20 30 40 50 60 70 80 90 100 110 120-2
-1
0
1
2IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120
-1.5
-1
-0.5
0
0.5
1
1.5
residue
Residue
To be an IMFNo. of maxima and minima = No. of zero crossings
orNo. of maxima and minima – No. of zero crossings = ±1
METHODOLOGY
26
EEG Online Database
27
http://epileptologie-bonn.de/cms Five data sets One hundred EEG records in each set Sampling frequency: 173.61 samples/sec Length of each record: 4097 samples Available with artifacts omitted
Healthy group Five healthy persons(awaken and relaxed) A:- eyes opened B:- eyes closed
Epileptic group Five patients C:-Inter-ictal (hippocampal formation) D:-Inter-ictal (epileptogenic zone) E:- Ictal (epileptogenic zone)
28
29
EEG samples
0 1 2 3 4 5 6 7 8 9 10-200
0
200
v
olts
Set B
0 1 2 3 4 5 6 7 8 9 10-200
0
200
v
olts
Set C
0 1 2 3 4 5 6 7 8 9 10-200
0
200
v
olts
Set D
0 1 2 3 4 5 6 7 8 9 10
-1000-500
0500
time (second)
v
olts
Set E
0 1 2 3 4 5 6 7 8 9 10
-100
0
100
v
olts
Set A
Feature Extraction Procedure
30
EEG
Low pass filter 9 Level EMD
Stat. feature
Chaotic feature
Stat. feature
Chaotic feature
Segmentation
Segmentation
Classification Procedure
31
Stat. Feature
Chaotic Feature Classifier
Classifier
Stat. Feature
Statistical Features
32
-2000 -1000 0 1000 20000
500
1000
1500
2000
2500Healthy EEG
Voltage (v)
Rel
ativ
e fre
quen
cy
-2000 -1000 0 1000 20000
500
1000
1500
2000
2500Inter-ictal EEG
Voltage (v)
Rel
ativ
e fre
quen
cy
-2000 -1000 0 1000 20000
500
1000
1500
2000
2500Ictal EEG
Voltage (v)
Rel
ativ
e fre
quen
cy
(a) (b) (c)
Statistical Features
Average level of dispersion:- Variance Average level of asymmetry:- Skewness Average level of peakedness:-Kurtosis
33
Statistical Features (Observation)
34
Preliminary observation
Quite large variance for ictal EEGs compared to non-ictal ones
Rapid decrease of variance from IMF1 to IMF9 In terms of variance, skewness and kurtosis – larger
difference among five sets in EMD domain
One-way ANOVA (99% confidence interval)
Variance:- Band-limited EEGs and all IMFs Skewness:- Band-limited EEGs and first two IMFs Kurtosis:- Band-limited EEGs and first four IMFs
Chaos
Apparent disorder of a non-linear system Highly sensitive to initial condition Phase-space plot demonstrates chaotic
behavior (if any) of time-series data.
35
Phase – Space Portrait
36
-200-150
-100
-50
0
50
100
150 -200
-150
-100
-50
0
50
100
150-200
-150
-100
-50
0
50
100
150
volts
Phase Space Plot of Healthy EEG
volts
vo
lts
Phase – Space Portrait
37
-400
-200
0
200
400
600
800
1000
1200
-400
-200
0
200
400
600
800
1000
1200-500
0
500
1000
1500
volts
Phase Space Plot of Inter-ictal EEG
volts
vo
lts
Phase – Space Portrait
38
-1000-800
-600-400
-2000
200400
600800
1000
-1000
-500
0
500
1000-1000
-800
-600
-400
-200
0
200
400
600
800
1000
volts
Phase Space Plot of Ictal EEG
volts
vo
lts
Chaotic Features
39
Chaoticity:- Largest Lyapunov Exponent (LLE)
Complexity:- Correlation Dimension (CD)
Level of disorder:- Approximate Entropy (ApEn)
40
Chaotic Features (Observation)Preliminary observation
No specific trend of LLE and CD values for Band-limited EEGs
Gradual decrease of LLE values from Set A to E for IMF1 and IMF2
Gradual decrease of CD values from Set A to E for IMF1 – IMF4
Gradual decrease of ApEn values from Set A to E for Band-limited EEGs, IMF1 and IMF2
For all IMFs, mean_LLE (Set A) > mean_LLE (Set E)mean_CD (Set A) > mean_CD (Set E)mean_ApEn (Set A) > mean_ApEn (Set E)
41
One-way ANOVA (99% confidence interval)
LLE:- Band-limited EEGs and first four IMFs CD:- First four IMFs ApEn:- Band-limited EEGs and first five IMFs
Chaotic Features (Observation)
ANN-based Classifier
42
Two-layer feed-forward network Hyperbolic tangent sigmoid transfer function Back-propagation training
Problem Formulation
43
Input Feature
SetClassifier
Healthy
Ictal Epilepsy Diagnosis
Input Feature
SetClassifier
Ictal
Inter-ictal
Seizure Detection
Problem Formulation
44
Input Feature
SetClassifier
Ictal
Inter-ictal
Detection and Diagnosis
Healthy
Diagnosis A, E
Detection D, E
Detection and Diagnosis A, D, E
Others (A,B,C,D), E (A,B), (C,D), E
RESULTS&
DISCUSSION
45
Classification Performance
46
Target ClassO
utpu
t Cla
ss
+ –
+
–
TP
TNFN
FP
Confusion Matrix
)/()()/()/(
FPTNFNTPTNTPAccuracyFPTNTNySpecificitFNTPTPySensitivit
Performance (3 Stat. Features)Detection (D, E) Better performance in EMD domain 100% accuracy for IMF3 and IMF4Diagnosis (A, E) Better performance in EMD domain 100% accuracy for IMF2 and IMF3Detection and diagnosis (A, D, E) Better performance in EMD domain 100% accuracy for IMF3
47
Performance (3 Stat. +3 Chaotic Features)
48
Detection (D, E) Better performance in EMD domain 100% accuracy for IMF3 and IMF4 Maximum accuracy reported to date:98.74% (FT, ApEn, AR modeling)Diagnosis (A, E) Better performance in EMD domain 100% accuracy for IMF2 and IMF3 Reported same performance using TFADetection and diagnosis (A, D, E) Better performance in EMD domain 100% accuracy for IMF3 Reported 99.28% accuracy using TFA
Stat. features – Maximum accuracy
Addition of chaotic features – similar performance
EMD domain statistical features play major role in characterizing EEGs
Explore the use of statistical modeling
49
Statistical Modeling
50
Sample data
Empirical pdf
Fitting with standard pdf
Estimated parameter
NIG and Stable pdfMoment based estimationEstimation requires large samples
Goodness-of-fit :- KLD, KS statistics
Statistical Modeling
51
0 5 10 15 20-200
-150
-100
-50
0
50
100
150
200
-200 -150 -100 -50 0 50 100 150 2000
0.002
0.004
0.006
0.008
0.01
0.012
Voltage label ( v)
Pro
babi
lity
Empirical pdf
-200 -150 -100 -50 0 50 100 150 2000
20
40
60
80
100
120
140
160
180
Voltage label ( v)
Rel
ativ
e fre
quen
cy
Histogram plot
52
Statistical Modeling
-2000 -1500 -1000 -500 0 500 1000 15000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6x 10
-3 Probability Density Function
Voltage ( v)
Pro
babi
lity
Den
sity
EmpiricalStable FitNIG Fit
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1P-P Plot
Probability
Pro
babi
lity
EmpiricalStable FitNIG Fit
Performance – Stat. Modeling
Band-limited EEGs – better fitting than IMFs
Stable pdf – better than NIG pdf
53
CONCLUSION
54
EMD domain statistics – better for detection and diagnosis
Less computational requirement Faster Less number of features ANN – optimum classifier Statistical model – promising method
55
FUTURERESEARCH
56
Extensive analysis using Stable pdf Maximum likelihood estimation Classification of visually inseparable EEGs Neo-natal seizure detection Sleep apnea detection Hypnosis detection Brain-computer-interfacing (BCI)
57
THANKYOU
58
QUESTIONS&
ANSWER
59