Scribd
Upload
40 views

Model Based Spectral Analysis

The document discusses model-based spectral analysis of biomedical signals like EEG. It describes autoregressive (AR) modeling and three methods for estimating AR model parameters: autocorrelation, modified covariance, and Burg's method. It also discusses calculating spectral parameters from the AR model like power, peak frequency, and bandwidth. Finally, it covers approaches for EEG segmentation including using the periodogram and whitening the signal with an AR filter to detect changes.

Uploaded by

Sathish Narayana Rao
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
40 views

Model Based Spectral Analysis

The document discusses model-based spectral analysis of biomedical signals like EEG. It describes autoregressive (AR) modeling and three methods for estimating AR model parameters: autocorrelation, modified covariance, and Burg's method. It also discusses calculating spectral parameters from the AR model like power, peak frequency, and bandwidth. Finally, it covers approaches for EEG segmentation including using the periodogram and whitening the signal with an AR filter to detect changes.

Uploaded by

Sathish Narayana Rao
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 26

T-61.

181 – Biomedical Signal


Processing

Chapters 3.4 - 3.5.2

14.10.2004
Overview
 Model-based spectral estimation
 Three methods in more detail
 Performance and design patterns
 Spectral parameters
 EEG segmentation
 Periodogram and AR-based approaches
Model-based spectral analysis
 Linear stochastic model
 Autoregressive (AR) model
 Linear prediction

x(n)  a1 x(n  1)  ...  a p x(n  p)  v(n)


1 1
H p ( z)  
Ap ( z ) 1  a1z 1 ... a p z  p
Prediction error filter
 Estimation of parameters based on
minimization of prediction error ep
variance
p
e p ( n )  x ( n)   ak x ( n  k )
k 1

 e2  E[e 2p (n)]
Estimation of model
parameters
 Parameter estimation process critical for
the successful use of an AR model
 Three methods presented
 Autocorrelation/covariance method
 Modified covariance method
 Burg’s method
 The actual model is the same for all
methods
Autocorrelation/covariance
method
 Straightforward minimization of error
variance
e p (n)  aTp ~
x p ( n)
~
 e2  E[e 2p (n)]  aTp R x a p

 Linear equations solved with Lagrange


multipliers (constraint apTi=1)
~
R x a p   e2 i
Levinson-Durbin recursion
 Recursive method for solving
parameters
 Exploits symmetry and Toeplitz
properties of the correlation matrix
 Avoids matrix inversion
 Parameters fully estimated at each
recursion step
Data matrix
 The correlation matrix can be directly
estimated with data matrices
~ ~ ~
R x  XTp X p
 In covariance method the data matrix does
not include zero padding, but the resulting
matrix is not Toeplitz
 In autocorrelation method the data matrix is
zero-padded
Data matrices in detail
Modified covariance method
 Minimization of both backward and
forward error variances
 Parameters from forward and backward
predictors are the same
 Correlation matrix estimate not Toeplitz
so the forward and backward estimates
will differ from each other
~
(R x  R x )a p   e2 i
Burg’s method
 Based on intensive use of Levinson-
Durbin recursion and minimization of
forward and backward errors
 Prediction error filter formed from a
lattice structure
Burg’s method recursion steps
Performance and design
parameters
 Choosing parameter estimation method
 Two latter methods preferred over the first
 Modified covariance method
 no line splitting
 might be unstable
 Burg’s method
 guaranteed to be stable
 line splitting
 Both methods dependant on initial phase
Selecting model order
 Model order affects results significantly
 A low order results in overly smooth spectrum
 A high order produces spikes in spectrum
 Several criteria for finding model order
 Akaike information criterion (AIC)
 Minimum description length (MDL)
 Also other criteria exist
 Spectral peak count gives a lower limit
Sampling rate
 Sampling rate influences AR parameter
estimates and model order
 Higher sampling rate results in higher
resolution in correlation matrix
 Higher model order needed for higher
sampling rate
Spectral parameters
 Power, peak frequency and bandwidth
 Complex power spectrum
 v2  v2
S x ( z)  1
 p
A( z ) A( z )
 (1
j 1
 d j z 1
)(1  d *
j z)

 Poles have a complex conjugate pair


d 2i  d 2*i 1
Partial fraction expansion
 Assumption of even-valued model order
 Divide the transfer function H(z) into
second-order transfer functions Hi(z)
p/2
H ( z)   H i ( z)
i 1

 No overlap between transfer functions


p/2 p/2
1
S x ( z )  H ( z ) H ( z )  
2
v
2
v  H ( z)H ( z
i 1
i i
1
)   S xi ( z )
i 1
Partial fraction expansion,
example
Power, frequency and
bandwidth
Pi  rxi (0)
Pi
Pi 
'

rx (0)
(d i )
i  arctan( )
( d i )
1  ri 2
i  arccos( cos i )
2ri
i  2(1  ri )
EEG segmentation
 Assumption of stationarity does not
hold for long time intervals
 Segmentation can be done manually or
with segmentation methods
 Automated segmentation helpful in
identifying important changes in signal
EEG segmentation principles
 A reference window and a test window
 Dissimilarity measure
 Segment boundary where dissimilarity
exceeds a predefined threshold
Design aspects
 Activity should be stationary for at least
a second
 Transient waveforms should be eliminated
 Changes should be abrupt to be
detected
 Backtracking may be needed
 Performance should be studied in
theoretical terms and with simulations
The periodogram approach
 Calculate a running periodogram from
test and reference window
 Dissimilarity defined as normalized
squared spectral error
 Can be implemented in time domain
N

 x
( r ( k , n )  rx ( k ,0 )) 2

1 ( n )  k  N
rx (0, n)rx (0,0)
The whitening approach
 Based on AR model
 Linear predictor filter “whitens” signal
 When the spectral characteristics change, the
output is no longer a white process
 Dissimilarity defined similarly to periodogram
approach
 The normalization factor differs
 Can also be calculated in time domain
Dissimilarity measure for
whitening approach
N t 1
1 e 2 (n  k )
 3 ( n) 
N

k 0
(
re (0,0)
 1)

 Dissimilarity measure asymmetric


 Can be improved by including a reverse
test by adding the prediction error also
from reference window (clinical value
not established)
1 Nt
et2 (n  k ) 1 Nr
er2 (k )
 4 ( n) 
Nt

k 1
(
rer (0,0)
 1) 
Nr
 (
k 1 ret (0, n)
 1)
Summary
 Model-based spectral analysis
 Stochastic modeling of the signal
 Is the signal an AR process?
 Spectral parameters
 Quantitative information about the
spectrum
 EEG segmentation
 Detect changes in signal

You might also like