Multichannel Analysis of the Newborn EEG Data

1
Multichannel Analysis of the Newborn EEG Data
Vaclav Gerla, Lenka Lhotska, Member, IEEE,
Vladimir Krajca, Karel Paul Czech
Technical University - Department of Cybernetics,
Prague - Czech Republic University Hospital
Na Bulovce, Prague - Czech Republic Care of
Mother and Child, Prague - Czech Republic
http//gerstner.felk.cvut.czgerlav_at_fel.cvut.cz
2
Our Research Purpose
EEG, ECG, EOG, EMG, PNG
Biological Signals
Mainly FFT/Wavelets
Feature Extraction / Selection
Classifier 1
Various type of classifiers Linear Models,
Neural Networks, Kernel Methods, Mixture Models,
Classifier 2

Optimalization
Classifier N
Weighted Average, Bagging, Boosting,Shafer
approach, Fuzzy Integral, BKS
Classifiers Combining
Visualisation in all stages of this process
Visualisation
We solve problem of feature extraction and we
compare various classifiers in this study
3
Motivation, Used Data

• Motivation, approach usability
• online monitoring
• estimation of the newborn brain maturity
• In this study we use data
• from 12 infants // 3 hours for each
• provided by the Institute for Care of Mother
  and Child in Prague
• Data are evaluated and scored by expert into 4
  stages
• quiet sleep
• active sleep
• wake
• movement artefact

proportion of these states is a significant
indicator in clinical practice!
4
System Structure
8 features
PSD (band 0.5-3Hz)
EEG, 8 channels
measure of regularity
PNG (respiration)
features centering Principal Component
Analysis (12 features 3 features)
beat frequency
ECG
EOG
PSD (1-2Hz)
EMG
standart deviation
F1
F2
F3
learning by EM
nearest neighbour
cluster analysis
decision rules
HMM
5
Segmentation
EEG
6
EEG Feature Extraction
- classification obtained by doctor- record
length 85 minutes
- features based on PSD- compute for each EEG
channel- delta band is shown here (0.5 to 3Hz)-
for subsequent processing we use these 8
characteristics
- simple classification procedure example- used
EEG signal only- based on proportion between
activities in the different EEG channels
(e.g.T3T4/C3C4)
7
EEG Feature Extraction
- PSD for other newborns signal- blue color
minimum red color maximum- maximum is in
central electrodes (C3, C4)
8
Regularity of Respiration Curve
- We utilize the strong regularity in quite
sleep gt autocorrelation analysis- clear
difference in the magnitude of the second peak in
the autocorrelation function- we use average
breath duration for second peak position
estimation
9
Regularity of Respiration Curve
- characteristics for other newborns- it is no
possible find one value for classification
threshold - but it is good for doctors (as
additional information )
10
Eye Movements
- we detect eye movements- derived from EOG
signal Algorithm 1. filter signal to freq.
band 1-2Hz 2. compute STDs in small windows
Utilized fact In the quiet sleep there should
not be any eye movements!
11
EMG Activity
- obtained from chin EMG signal- computed STD of
this signal- feature useful for movement
artifact detection- we compute mean value for
small window (removing peaks) and than we find
maximum for bigger windows (trend enforcement)
Utilized fact Large majority of movement
artifacts are present at EMG signal
(characterized by the very high amplitude)
12
EMG Activity
- muscles activity for other newborns- not
present in quiet sleep
13
Heart Rate
- derived from ECG- used standard method for QRS
position detection based on first derivation-
we detect maximum of R-peakThe amplitude and
the regularity of heartrate is changed during
sleep!
14
Heart Rate
- heart rate characteristics for other newborns-
slow changes are visible- heart rate is lower in
quiet sleep
15
Principal Component Analysis

• reduce the number of dimensions without
  significant loss of information
• original features are very correlated -gt
  PCA saves classification time

PCA
16
Hidden Markov Models

• in our case, HMMs allow us to describe
  relations between all features and hidden states
  (all sleep stages)
• we use the EM algorithm for finding the
  maximum-likelihood estimate of the parameters of
  HMMs
• choise of initial model is crucial - we compute
  it from the training data set

mutual relations between individual hidden states
17
Results
Accuracy of classification
1. We used all data from 12 newborns and
cross-validation (10 group)
2. We used data from 11 newborns for learning and
data from remaining one newborn for testing. This
procedure we repeated for all newborns and
computed mean value.
18
Conclusion

• our final accuracy obtained was about 70 on
  unknown data set compared with physician
  (evalution accuracy of physician is about 80)
• very illustrative is to show final decision
  together with all described characteristics (we
  can see significant trends during sleep)
• during automated classification we have problem
  with clear separation of stages wake and active
  sleep. Now we try to find hidden information
  enabling this separation
• our designed technique can be applicable to
  other similar problem in medicine as well

19
Future Work

• in our further research we plan to develop
  methods for quantification that can help in
  evaluation of newborns brain maturity
• we expected increasing of accuracy and
  robustness by the combining all described
  classifiers. We plan use methods as bagging and
  boosting

• we plan to use similar methods for
  classification of sleep in adults

• we have developed hardware solution for on-line
  measuring of EEG (now we concentrate on the pda
  based analysis methods)

20
Thank you for your Attention