The aim the project is to analyse non real time EEG (Electroencephalogram) signal using different mathematical models in Matlab to predict abnormal derivation of the signal applying frequency spectral analysis for linear, continuous or discrete input

1
Introduction
Scanning and Detection of EEG Diseases Using
Medical Signal Processing
1

• The aim the project is to analyse non real time
  EEG (Electroencephalogram) signal using different
  mathematical models in Matlab to predict abnormal
  derivation of the signal applying frequency
  spectral analysis for linear, continuous or
  discrete input data signal. This will involve a
  filtering pre-processing stage, Short Time
  Fourier Transform, DFT, FFT, AR Model,
  Sonification and Hidden Markov Model (HMM) for
  more that one signal with a further application
  in Bayes Networks Classifier.

Objectives

• The project will study new techniques for the
  analysis of EEG and the automated diagnostic of
  the pathologies.
• Data will be analysed using AR model because this
  technique will study information extraction from
  signals that are a- periodic, noisy, intermittent
  or transient from a tiny signal, which contain
  very small amplitude and period .
• Sonification of the EEG data is applied to obtain
  an acoustic representation of the signal in a
  spectral form. The sonification technique will
  convert the spectrogram frequencies of the EEG
  data in audible sound to detect the disease.
• Hidden Markov Model (HMM) will process different
  EEG data as stochastic sequences of events.
• EEG data will be imported by Matlab and the model
  is applied in a selected normal and abnormal
  signal as Epilepsy, Arrhythmia or whatever EEG
  supplied data .

2
Methodology
2

• The system has analysed two different data sets
  from the next sources The 1st data source
  contains normal EEG data from Colorado State
  University The 2nd data source contains Epilepsy
  EEG data from Bonn University (Germany).
• EEG Data is provided in mat file or txt file.
  Matlab will give the option to create scripts for
  the models using the DSP, System Identification,
  Hidden Markov Model (HMM), Wavelet Transform and
  Neural Networks toolbox,.
• Feature Extraction EEG signal will be
  pre-processed to eliminate the noise using the
  Band Pass filter Butterworth IIR because the 1st
  data set contains noise as row signal. It can
  affect to the next applied models, but the wrong
  results affects mainly to the periodogram.
• AR (Autoregressive) Model will study the
  behaviour of the EEG signal coefficients for
  large or small frequency samples in linear form.
  ARburg model is applied for small EEG data
  windows and frequency samples.
• Sonification model will analyse the spectrum of
  the signal by differential sonification and Short
  Time Fourier Transform (STFT) to find the
  harmonics and lobe bands. Frequencies generates
  audible tones (5 to 90 Hz).
• Hidden Markov Model (HMM) analyses data to detect
  the diseases by observation of the input classes
  or sequences. Also HMM classifies it by events in
  a Gaussian 2D of each state of the signal. Then
  the signal will contain a sequence of events
  called Markov Chain with Gaussian densities.
  Bayesian Classification estimates the optimal
  sequence by Viterbi Algorithm.

3
AR Model
Sonification
3
Normal EEG Data, AR 9th , 10th and 11th window.
Blinking Eyes
Epilepsy EEG Data, AR 7th , 8th and 9th window.
Normal Spectrogram EEG Data 9th, 10th and 11th
window.
Epilepsy Spectrogram EEG Data 7th, 8th and 9th
window.
Hidden Markov Model (HMM)
Periodogram EEG C3 (noisy line) channel
Periodogram EEG Epilepsy 7th window channel
Model
Accuracy () p-value
EEG Data HMM-1 Must be 100
Must be 1.0
EEG Data HMM-2 Must be 100
Must be 1.0
4
Results
4

• AR model estimates the arburg coefficients from
  normal EEG signal and Epilepsy signal..
• Normal EEG Data linear vector.
• Epilepsy EEG Data logarithmic curve
  vector with an optimal point to show the critical
  state in the seizure.
• Sonification
• 1. The Probability Density Estimation
  calculates three gaussian kernel bandwidth
  approximations (default widths).
• Normal EEG Data Gaussian widths almost
  matched.
• Epilepsy EEG Data Mismatch gaussian
  widths.
• 2. The spectrograms show small amplitude
  values for light colours and high amplitude
  values for dark colours in the Short Time Fourier
  Transform. The intensity of the frequency colours
  give the harmonics of the pattern plotted.
• 3. Spectral sonification is audible to
  human ear (5Hz to 90Hz). The amplidude of the EEG
  signal changes the tone range.
• Hidden Markov Model (HMM) EEG values have to be
  -5 to 5 to avoid mismatches between data and
  initial random process. EEG signals are low
  correlated, except sleep stages. AR coefficients
  (EEG signal) are trained in two models with
  higher (log-) likelihood value. HMM1 and HMM2
  models are compared in the table to show the
  classification accuracies and the intervals with
  the standard deviation.

Future Work

• Implement Hidden Markov Model (HMM) using the
  Factorial Markov Model (FMM) and Boyen-Kollen
  algorithm for a Bayes Network Classifier.
• Classification using Neural Network Classifier.
• EEG analysis using Wavelet Transform and
  classification of the Wavelet Feature Extraction.

Luis Acevedo MSc Embedded Systems
(2004)Supervisor Dr. Yvan Petillot