Prediction of Epileptic Seizures PhD Conversion Seminar

1
Prediction of Epileptic Seizures PhD Conversion
Seminar
Elma OSullivan-Greene Life Sciences, NICTA
VRL Dept. Electrical Electronic Engineering,
The University of Melbourne elmao_at_ee.unimelb.edu.a
u
Supervisors Prof. Iven Mareels Dr.
Levin Kuhlmann A/Prof. Anthony Burkitt
Dr. Chung-Yao Kao
2
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

3
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

4
Epilepsy a disorder of the brain

• Epilepsy is a neurological disorder
• Characterised by recurrent seizures
• Associated with abnormally excessive or
  synchronous neuronal activity in the brain
• Most common serious neurological condition
• Prevalence of epilepsy varies across geographical
  regions within the range of 0.5 to 4 of the
  total population (WHO)
• Current Treatment
• AED (Antiepileptic Drugs) - undesirable
  side-effects
• Surgical removal of the epileptic brain tissue

5
Motivation For Seizure Prediction

• The ability to predict seizures would have a
  profound impact on the quality of life of
  epilepsy suffers.
• Our proposed solution
• An Implantable device incorporating
• seizure prediction
• short-term electric stimulation treatment for
  seizure prevention
• Continuous electric stimulation is in use,
  and shows good results in
  many patients (unknown side
  effects for long term use)
• No robust seizure prediction algorithm has been
  published to date

6
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

7
Data Source Electroencephalography (EEG)

• Recordings of the fluctuating electric fields of
  the brain
• Electric fields due to ionic currents in the
  extra cellular fluid
• Neurons (nerve cells) choose when to fire
  impulses based on this ionic current information

8
Data Source Electroencephalography (EEG)

• Recordings of the fluctuating electric fields of
  the brain

• Scalp EEG data

• Intracranial EEG data

9
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

10
Can Seizures Be Predicted?

• Evidence for a definable pre-ictal (pre-seizure)
  period
• Clinically undisputed indicative systematic
  changes are present in some patients prior to
  seizure onset
• Mood changes, nausea, headache
• Several signal processing studies argue that a
  pre-ictal state can be defined based on
• Measures of synchronisation between EEG channels
• Non-linear dynamics measures

11
Current Prediction Approaches

• Linear Approaches
• Spectral analysis
• Linear Modelling
• Energy measures
• Minimal success brain function nonlinear?
• Nonlinear Approaches
• Based on state space reconstruction
• Dimension
• Lyapunov Exponents
• Entropy
• Minimal success initial promising results failed
  to be reproduced with other data sets

12
State Space Reconstruction/ Delay Embedding
N at least O(1015)
13
State Space Reconstruction/ Delay Embedding
Combine to reconstruct an N-dimensional system
14
Limitations of Delay Reconstruction

• The original framework (Takens/ Aeyels) for
  delay reconstruction requires
• Stationarity of the dataset
• Noise free data set
• A time series from an autonomous dynamical system
• Low dimensionality of underlying dynamical system
• However the EEG is ultimately an unsuitable
  signal for this framework
• Highly non-stationary data set
• High levels of measurement noise in EEG
  recordings (artefact)
• The brain is not an autonomous system (brain
  processes external inputs)
• No conclusive evidence that the brain/ epileptic
  events are low dimensional

15
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

16
Work Completed and in Progress

• Modification of the EEG signal for delay
  reconstruction
• Addressing the noise limitation
• Taking the difference between 2 closely spaced
  intracranial electrodes

• Cancel common mode input from far away dynamical
  subsystems (Stark)
• Representation of local dynamics
• Significant reduction in common mode artefact (50
  Hz mains pick-up)

• Consider the Brain as
• spatially distributed system
• interacting distinct
  local
  subsystems

17
Work Completed and in Progress

• Modification of the EEG signal for delay
  reconstruction
• Addressing the low dimensionality limitation
• Hypothesis The brain is lower-dimensional during
  a seizure
• Perhaps there is enough stationary data in the
  period just prior to a seizure to warrant a
  reconstruction

18
An existing seizure prediction algorithm

• Dynamical Similarity Index (DSI)
• Le Van Quyen (1999)
• Creates templates of brain dynamics from delay
  reconstruction of EEG data
• Seizure anticipation state declared for large
  sustained deviations of dynamic template from
  reference (far from seizure)

19
Work Completed and in Progress

• Application of modified EEG signal to DSI
  algorithm
• Reference template from pre-ictal data
• (low-dimensional/stationarity considerations)
• EEG signal used difference between 2 closely
  spaced intracranial electrodes
• (noise consideration)
• Preliminary results
• Sensitivity 25-100 across 3 patients
• False Positive rate 1-6.6 FP/hr across 3 patients

20
Work Completed and in Progress

• No major improvement seen with preliminary
  results over original DSI algorithm
• Why?
• Pre-ictal low dimensionality of underlying system
  is an unproven hypothesis
• Other noise muscle artefact, cardiac artefact
• Conclusion
• Future prediction methods should concentrate on
  non-delay-reconstruction based methods

21
Talk Outline

• Epilepsy a disorder of the brain
• Data available for engineering analysis
• Current approaches to epileptic seizure
  prediction, and their limitations
• Work completed and in progress
• Proposed avenues for project

22
Project Proposal

• Nonlinear System analysis without reconstruction
• Data-driven pathway

Brain System Unknown state space system,
F xk1F (xk , ?k)
Measured EEG Data Represented by the function,
H zkH (xk , ?k)
Epilepsy Prediction Represented by the function,
G yk1G (xk , ?k)
?
23
Project Proposal - Entropy via Data Compression

• An entropy measure as a prediction candidate
• Low-dimensional object indicative of underlying
  brain state
• Entropy, as measured in the brain, can be viewed
  as
• a measure of how chaotic the brain system is
• a measure of information transfer in the brain

24
Entropy via Data Compression Techniques

• Instead of computing entropy via delay
  reconstruction.
• Estimating entropy via Data-Compression
  Techniques
• Markov Model
• Context-Tree Model
• Model based on Independent Component Analysis
  (ICA)
• Let observed time-series data (EEG) be an element
  of a finite alphabet of symbols
• Advantages of this approach
• More robust in the presence of noise
• Does not require stationarity of the data set
• Can be applied to High Dimensional Systems

25
Entropy Estimation from a Markov Model

• Markov model
• Estimates future symbols based on k-past past
  symbols
• Symbolic time series analysis

26
Entropy Estimation from a Weighted Context Tree

• Weighted Context tree
• Estimates future symbols based on k-past past
  symbols
• Each node or context contains information of
  symbol history
• Automated recursive weight probability associated
  with each context
• Contexts automatically discarded on basis of
  improved performance
• Entropy h L / N LSource Code
  length NTime

27
Entropy Estimation from an ICA based model

• Measured EEG Channels x A s
• Find the transformation of the data W A-1 such
  that the coding lengths of the components are
  minimised
• Non-linear independent component methods
• Using several EEG channels spatial information

Statistically Independent components
x f( s ) y h( x )
28
Seizure Prediction Proposal

• Have discussed Entropy as a seizure prediction
  candidate as estimated from data compression
  techniques.
• Next An alternative probabilistic approach to
  data-based seizure prediction

29
Seizure Prediction Decision Markov Process

• Motivation for a statistical decision model
• Dynamical systems representations of the
  epileptic brain

• Bifurcation Phenomenon prediction by tracking
  the trajectory of bifurcation parameter, µ, over
  time
• Bifurcation part of thalmo-cortical brain
  model, Robinson (2003)

• Probabilistic Transitions between two chaotic
  attractors
• Normal Epileptic
• Phase portrait of computer model of brains
  thalmo-cortical network, Lopes Da Silva (2003)

30
Seizure Prediction Decision Markov Process

• 3 state model
• Transition probabilities tij assigned through
  analysis of EEG data
• Potential for intervention applications control
  input to minimise the transition to seizure state

31
Conclusion Research Proposal

• Proposed Avenues for Seizure Prediction
• Entropy as estimated from data compression
  techniques
• Markov Process
• Context Tree
• Independent Component Analysis
• Decision Markov Process
• Potential for the theoretical expansion of
  dynamical system time-series analysis
• for the application of real world biological
  data

32

• Thank you for your attention
• Questions?
• elmao_at_ee.unimelb.edu.au