An EEG Based Computational Model for Seizure Detetcion

This presentation consists of detailed analysis of EEG signals , preictal time prediction

1. An EEG-Based Computational Model
for Epileptic Seizure Detection and
Preictal Time Prediction
Presented by
D.SHYAM SUNDAR
(3122213002098)

2. Problem Statement
Developanaccurateandtimelycomputationalmodel
usingEEGdatatoclassifythekindsofseizuresand
predictepilepticseizuresinadvance,improving
interventionandoptimizingepilepsymanagement.

3. Outline
•Objective
•Motivation
•LiteratureSurvey
•Introduction
•DatasetDescription
•Work
•Results
•Conclusion
•References

4. Objective
•AccurateSeizurePrediction:Developamodelto
predictepilepticseizureswithhighaccuracyusing
EEGdata.
•TimelyDetection:Identifythepreictalphasetoenable
earlyinterventionandimprovepatientsafety.
•FeatureExtraction:Utilizeadvancedtechniquesfor
extractingrelevantfeaturesfromEEGsignalsto
enhancepredictionreliability.
•ClinicalIntegration:Createamodelthatcanbeeasily
integratedintoclinicalpracticeforeffectiveepilepsy
management.

5. Motivation
•ImproveSafety:Earlyseizurepredictionreducesthe
riskofinjuryforepilepsypatients.
•BetterClinicalOutcomes:Enhancetreatmentand
managementthroughreliableseizurepredictions.
•AdvanceUnderstanding:Gaininsightsintopreictal
brainactivitypatternsandseizuremechanisms.
•OvercomeLimitations:Addresstheaccuracyand
invasivenessissuesofcurrentepilepsymanagement
technologies.

6. Literature Survey
InferenceAuthors Title of the PaperS.No
FeatureExtraction
fromEEGsignals
intimedomainand
frequencydomain
K. Kannadasan
B. Shameedha
Begum
V. Sridevi
AnEEG-BasedComputational
ModelforDecodingEmotional
Intelligence,Personality,and
Emotions
1
Earlydetectionof
epilepticseizures
bydetectingthe
spikesintheEEG
waveform
Khakon Das
Partha Pratim Roy
Atri Chatterjee
Shankar Prasad
Sha
Seizurepredictionbythedetection
ofEEGwaveformfromthepre-
ictalphaseofEEGSignals
2
Seizuredetection
basedonthe
frequencybandsin
thebrain
A.M. Chan
F.T. Sun
E.H. Boto
Automatedseizureonsetdetection
foraccurateonsettime
determinationinintracranial
EEG
3
Seizuredetectionin
thepreictalphase
S.S.P. Kumar
L. Ajitha
Early detection of Epilepsy using
EEG Signals
4

7. Introduction
•Epilepsyisaneurologicaldisorderwithrecurrent
seizuresthatsignificantlyaffectpatients'qualityoflife.
•Existingseizurepredictionmethodslackaccuracyand
timelydetection,limitingeffectiveintervention.
•Thisprojectfocusesonaccuratepreictaltime
prediction,providingearlywarningsforbetterseizure
management.
•Theprojectaimstodevelopacomputationalmodel
usingEEGdatatoimproveseizureprediction,enhance
patientsafety,andadvanceclinicalpractices.

8. Phases of Epilepsy

9. Dataset Description
•The Bonn EEG Dataset is a collection of EEG
(electroencephalogram) recordings used for neuroscience and
machine learning research.
•This dataset contains 5 signal files each with 4096 samples with
classes like normal, partial seizure, general seizure, focal seizure,
and myoclonic seizure.
•The sampling rate of the data was 173.61Hz for better
performance. In EEG recordings, sampling is necessary because
it allows for the digital processing and analysis of the signals.
•The application of a low-pass filter, with fc = 40 Hz is the first
step involved in processing to remove the noise and unwanted
components.

10. Work -Methodology
Global Architecture of the Proposed Model

11. Work –Data Visualisation
•TimeDomainVisualisation:Continuous-timeSignalPlotoftheEEG
Waveformfromeachclassofthedataset
•FrequencyDomainVisualisation:UsageofContinuousWaveletTransform
(CWT)forscalogramandShortTimeFourierTransform(STFT)for
spectrogramtorepresentthefrequencycontentofthesignal
Time Domain –Signal Plot Frequency Domain –Scalogram

12. Data Preprocessing
•Up-sampling: The data is interpolated (up-sampled) to 200 Hz to increase the
temporal resolution of the EEG signals.
•Down-sampling: Followed by interpolation, the data is decimated (down-
sampled) to 100 Hz to reduce the computational load and retain essential
information
•Filtering: A low-pass filter with a cut-off frequency of 40Hz is applied to
eliminate high frequencies
•Noise Removal: Usage of Haar wavelet for artifact removal techniques.
Filtered Signal

13. Data Analysis
•DescriptiveAnalysis:Themean,standarddeviation,median,
andpercentiles(25thand75th)arecalculatedtounderstandthe
centraltendencyandvariabilityofEEGsignals.
•EEGAnalysis:TheWelchmethodisusedtoanalyzefrequency
components,identifyingdominantfrequenciesandpower
distributionacrossdifferentbands.
•SpectralAnalysis:Generatebraintopographicmapstovisualize
powerdistributionacrossbrainregions,focusingonDelta,
Theta,Alpha,andBetabandsforseizurelocalization.

14. Cont..
Range of OperationFrequency band
0.5 -4 HzDelta Band
4 -8 HzTheta Band
8 -12 HzAlpha Band
13 -30 HzBeta Band
Table 1 Frequency bands operating ranges

15. Feature Extraction
•TimeDomain:FeatureslikeNormalizedFirstDifference(NFF),
HjorthMobility(HM),HjorthComplexity(HC),andHigherOrder
Crossings(Hoc)provideinformationonsignaldynamicsand
patterns.
•FrequencyDomain:DerivedfromPowerSpectralDensity(PSD)
oftheta,delta,alpha,andbetabandsusingtheWelchmethodto
captureenergywithinspecificfrequencyranges.
•Time-FrequencyDomain:Combinestimeandfrequencyaspects
withfeatureslikeRootMeanSquare(RMS)andRecursiveEnergy
Efficiency(REE)toquantifyenergydistributionusingContinuous
WaveletTransform(CWT)fordetailedsignalrepresentation.
•FeatureCompilation:Extractfeatures,organizethemintoa
structureddataset,savetheminCSVformat,andpreparefor
analysisandmachinelearningmodeltraining.

16. Cont…
Table 2 Time Domain Features
ExpressionParameter
Maximum amplitude / Mean amplitudeNormalized First difference (NFF)
(var(y’(t))/var(y(t))
0.5
Hjorth Mobility (HM)
mobility(y’(t))/mobility(y(t))Hjorth Complexity (HC)
Number of times a signal crosses its
mean value
Higher-order Crossings (HoC)
AftertheextractionofalltherequiredfeaturesintheTimedomain,theFrequency
domain,andtheTime-Frequencydomain,theextractedfeaturesarecompiledinto
afeaturesetandconvertedtoaCommaSeparatedValuesfile(.csvfile)forthe
classificationandpreictaltimeprediction.

17. Seizure Classification
•To classify the different kinds of seizures, machine learning
classification models like K-Nearest Neighbours(KNN), Support
Vector Machine(SVM), Logistic Regression(LR), Linear
Discriminant Analysis(LDA), Decision Tree(DT) and Naive
Bayes(NB) are used.
•The seizures are classified as Healthy (Normal –No seizure),
Focal seizures, General seizures, Partial Seizures, and Myoclonic
Seizures.
•Each kind of seizure is characterized by its unique features, EEG
patterns, location of origination in the brain, different kinds of
symptoms and neuronal activity in the brain.

18. SymptomsEEG PatternOriginType of Seizure
motor, sensory,
autonomic, or
psychic symptoms.
Focal epileptiform
discharges,
localized
abnormalities.
Specific area of the
brain
Partial Seizures
tonic-clonic
movements, brief
lapses in awareness,
sudden muscle tone
loss
Generalized spike-
and-wave,
polyspike-and-wave
discharges
Both hemispheres
from onsetGeneral Seizures
Similar to partial
seizures.
Focal spikes, sharp
waves, localized
slowing.
Specific area
(similar to Partial
Seizures)
Focal Seizures
Sudden , brief ,
involuntary muscle
jerks
Generalized,
bilateral,
synchronous,
polyspike –and –
wave discharges
Diffuse,
generalized
(specific
syndromes)
Myoclonic Seizures
Cont…

19. Cont…

20. Cont…

21. Preictal time prediction
•FeatureExtraction:ThescalogramfeaturesusingCWTand
bandpowerfeaturesfromPSDinspecificfrequencybands
(delta,theta,alpha,beta)areextractedfromtheEEGfiles.
•ThresholdDefinition:Specificthresholdsforeachfrequency
bandareestablishedtodifferentiatepreictalperiodsfrom
interictalandictalstates,identifyingsignificantchanges
indicativeofanimpendingseizure.
•RNNModel:TheextractedfeaturesarefedintoaRecurrent
NeuralNetwork(RNN)modeltocapturetemporal
dependenciesinEEGsignalsandpredictpreictaltimeby
learningpatternsassociatedwiththepreictalphase.
•SMSAlert:AnSMSalertisgeneratedandsenttothepatient’s
mobile,providinganadvancedwarningofanupcomingseizure
fortimelyinterventionandmanagement.

22. Results –ML Models
The performance metrics of the ML classification models
used for the classification of different kinds of seizures are
shown in the below table
F1 scoreRecallPrecisionAccuracyModel
0.870.960.800.81KNN
0.850.880.810.80LDA
0.820.800.840.80LR
0.741.000.680.74NB
0.840.920.770.78SVM
0.900.920.880.82DT

23. Results –RNN Model
•ExtractedscalogramandpowerbandfeaturesarefedintotheRNN
model.
•Adamoptimizerisemployedforefficienttrainingover50epochs.
•Thedatasetissplitinto70%training,15%validation,and15%
testingforrobustevaluation.
•-Modelachieves93%trainingaccuracyand89%validation
accuracy,indicatingeffectivelearningandgeneralization.
•Accuracyandlossfunctionplotsvisualizetrainingdynamics,
confirmingthemodel'scapabilityinpredictingpreictaltime.

24. Results -Training and Testing
Plot of Accuracy Curve Plot of Loss function Curve

25. Results –Alert Generation

26. Results –Alert Generation

27. Results –Clinical Validation
•Thepredictedpreictaltimeresultswerevalidatedbya
consultantneurosurgeon,onascaleof0-5where0represents
“Poorprediction”and5for“Prefectprediction”.
•Theaverageoftheresultswastakenandstatisticalanalysisof
theresultsisperformed.

28. Novelty of the work
NoveltyMethods usedTitle of the Paper
Focus on the classification of seizures and
partial preictal prediction
Machine Learning
classification
models
Earlydetectionof
EpilepsyusingEEG
signalsbySelvin
PradeepKumaret.al
Classification of preictal and interictal stages
of epilepsy
DCNN and
Bidirectional RNN
DeepLearning-based
ReliableEarly
EpilepticSeizure
PredictorbyHisham
Daoudet.al
Classification of normal and epileptic EEG in
the preictal stage
Feature extraction
and Pattern
matching
Epilepticseizure
predictionbythe
detectionofseizure
waveformfromthe
pre-ictalphaseof
EEGsignal
Multidomain feature extraction, classification
of seizure, and preictal time prediction and
alert generation
ML classifiers and
RNN with LSTM
Thiswork

29. Conclusion
•Thisstudypresentsacomprehensivemodelforepilepsydetectionand
preictaltimepredictionusingmachinelearning,leveragingfeaturesfrom
multipledomainsandanRNNforpreciseinterventiontiming.
•Thefuturevisionistodevelopawearabledevice,likeasmartwatch,to
predictseizures20minutesinadvance,enhancingepilepsymanagement
andpatientsafety.
Wristband wearable for epilepsy alert generation

30. SPECIAL THANKS
Dr . Nikunj Arunkumar Bhagat
Assistant Professor,
Department of Electrical Engineering,
Department of Biological Sciences and
Biosciences,
Indian Institute of Technology, Kanpur
Dr. Vighneshwar Ravishankar
Consultant Neurosurgeon,
Institute of Neuroscience,
Apollo Hospitals

31. References
•K.Kannadasan,J.Shukla,S.Veerasingam,B.S.Begum,andN.Ramasubramanian,
"AnEEG-BasedComputationalModelforDecodingEmotionalIntelligence,
Personality,andEmotions,"IEEETransactionsonInstrumentationand
Measurement,vol.73,Art.no.2505413,2024.
•A.M.Chan,F.T.Sun,E.H.Boto,andB.M.Wingeier,"Automatedseizureonset
detectionforaccurateonsettimedeterminationinintracranialEEG,"Clin.
Neurophysiol.,vol.119,no.11,pp.2572-2579,Nov.2008.doi:
10.1016/j.clinph.2008.08.025.
•E.Chesktor,K.Das,D.Daschakladar,P.P.Roy,A.Chatterjee,andS.P.Saha,
"Epilepticseizurepredictionbythedetectionofseizurewaveformfromthepre-ictal
phaseofEEGsignal,"BiomedicalSignalProcessingandControl,vol.57,p.
101720,2020.
•T.JhansiRaniandD.Kavitha,"AstudyonEEGsignalsforepilepticseizure
detectionusingmachinelearningclassifiers,"inProceedingsofthe6thInternational
ConferenceonCommunicationandElectronicsSystems(ICCES-2021),IEEE
XplorePartNumber:CFP21AWO-ART,ISBN:978-0-7381-1405-7.

32. Thank You!