Journal of Neuroscience Methods 196 (2011) 151158

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

journal homepage: www.elsevier.com/locate/jneumeth

Time-varying model identication for timefrequency feature extraction from

EEG data
Yang Li a,1 , Hua-Liang Wei a,2 , Stephen A. Billings a, , P.G. Sarrigiannis b
a
b

Department of Automatic Control and Systems Engineering, The University of Shefeld, Mapping Street, Shefeld S1 3JD, UK
Department of Clinical Neurophysiology, Shefeld Teaching Hospitals NHS Foundation Trust, Royal Hallamshire Hospital, Glossop Road, Shefeld S10 2JF, UK

a r t i c l e

i n f o

Article history:
Received 28 October 2010
Received in revised form
26 November 2010
Accepted 28 November 2010
Keywords:
TVARX model
Wavelet basis functions
Model structure detection
Recursive least squares (RLS)
Orthogonal least square (OLS)
Time-dependent spectra
EEG

a b s t r a c t
A novel modelling scheme that can be used to estimate and track time-varying properties of nonstationary signals is investigated. This scheme is based on a class of time-varying AutoRegressive with an
eXogenous input (TVARX) models where the associated time-varying parameters are represented by
multi-wavelet basis functions. The orthogonal least square (OLS) algorithm is then applied to rene the
model parameter estimates of the TVARX model. The main features of the multi-wavelet approach is that
it enables smooth trends to be tracked but also to capture sharp changes in the time-varying process
parameters. Simulation studies and applications to real EEG data show that the proposed algorithm can
provide important transient information on the inherent dynamics of nonstationary processes.
2010 Elsevier B.V. All rights reserved.

1. Introduction
Time-varying (TV) processes encountered in different engineering applications such as biomedical signal processing can be characterised by parametric representations (Wei and Billings, 2002;
Chon et al., 2005; Chen and Chowdhury, 2007; Wei et al., 2010).
Thus, the need to identify TV systems has naturally led to a growing interest in these areas. Parameter identication and modelling
is now established based on the AutoRegressive with an eXogenous
input (ARX) model. The ARX model, which can match the structure
of many real-world processes, is one of the most widely applied
linear dynamic models. The popularity and wide application of the
ARX model comes mainly from its easy-to-compute parameters
(Nells, 2001; Burke et al., 2005; Wei et al., 2009).
Many approaches have been proposed to identify time-varying
AutoRegressive with eXogenous input (TVARX) models. One of the
most popular techniques to deal with TV modelling problems is to
adopt an adaptive algorithm such as the Kalman lter and recursive least squares approaches (RLS) (Ljung and Gunnarsson, 1990;

Corresponding author. Tel.: +44 0114 222 5232; fax: +44 0114 222 5683.
E-mail addresses: coq08yl@shefeld.ac.uk (Y. Li),
w.hualiang@shefeld.ac.uk (H.-L. Wei), s.billings@shefeld.ac.uk (S.A. Billings),
Ptolemaios.Sarrigiannis@sth.nhs.uk (P.G. Sarrigiannis).
1
Tel.: +44 0114 222 5553; fax: +44 0114 222 5683.
2
Tel.: +44 0114 222 5198; fax: +44 0114 222 5683.
0165-0270/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jneumeth.2010.11.027

Chowdhury, 2000, 2003; Jiang and Chowdhury, 2005). Another

method for TV system identication is to expand the TV parameters
onto a linear or nonlinear combination of a set of basis functions. Consequently, the unknown time-varying parameters are
then converted into a set of constant coefcients of the basis functions. Hence, the initial TV identication and modelling problem
is simplied and becomes a deterministic regression selection and
parameter estimation problem.
The choice of basis functions can signicantly affect the performance of the parameter estimates. However, there is no guideline
on how to choose the appropriate basis functions for a specic
modelling problem. Conventionally, the basis functions have been
chosen to be polynomials (including Chebyshev and Legendre
types), prolate spheroidal sequences which are the best approximation to bandlimited functions (Niedzwiecki, 2000; Wei and Billings,
2002; Zou and Chon, 2004; Chon et al., 2005) and wavelets that
have a distinctive property of multi-resolution in both the time
and frequency domains (Cakrak and loughlin, 2001; Ansari et al.,
2005; Wei et al., 2010). In fact, each family of basis functions has its
own properties of accuracy and tractability, for example, polynomial and Fourier basis functions can work well for most smoothly
and slowly varying coefcients; wavelet basis functions, however,
perform well for time-varying coefcients of abrupt changes or
piecewise variations. Basis function expansion methods have been
widely applied to solve various engineering problems. For example,
a TVAR model can be expanded over a FourierBessel (FB) series to
constitute a feature vector for segmentation of the EEG signal, and

152

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

then to nd a simple model for the parametric representation of

EEG signals (Pachori and Sircar, 2008). A good choice of the basis
functions should allow abruptly or rapidly changing parameters to
be tracked.
Wavelets have distinctive approximation properties and are
well suited for approximating general nonstationary signals
(Mallat, 1989; Chui, 1992; Wei and Billings, 2002, 2004, 2006; Wei
et al., 2010), and thus have been successfully applied to many areas
including nonlinear signal processing and parametric identication (Tsatsanis and Giannakis, 1993; Adeli et al., 2003; Wei et al.,
2004; Jason et al., 2008; Ng and Raveendran, 2009). However, to
our knowledge, not much work has been done to exploit the inherent approximation properties of wavelets to identify TV coefcient
parameter estimation. The objective of this study is to present
a novel TVARX modelling approach, where the time-dependent
coefcients are expanded using a nite set of multi-wavelet basis
functions. Based on a multi-wavelet expansion scheme, a new
method for time-dependent parameter estimation is then proposed. The term multi-wavelet here has a twofold meaning. Firstly,
the TV coefcients of the ARX model are approximated using
several types of wavelet basis functions (i.e. the TV parameter
estimation involves multiple wavelets). Secondly, these wavelet
basis functions are combined in a form of multi-resolution wavelet
decomposition. The advantage of the proposed method, compared
with a method involving only a single type of wavelets, is that the
multi-wavelet expansion scheme is much more exible in that it
exploits the excellent properties of both non-smooth and smooth
wavelet basis functions and thus can effectively track both rapid
and slow variations of TV coefcients. In addition, the expansion
of TV parameters onto multi-wavelet basis functions is more accurate and effective for dealing with nonstationary signal modelling
than traditional power spectral estimation approaches and classical
time-invariant parameter models.
Brain signals involved in neuro-physiological techniques such
as EEG and MEG are inherently nonlinear and nonstationary
processes which exhibit complex dynamics. Conventional timedomain analysis methods such as FFT and power spectral density
(PSD) estimation methods, where it is assumed that signals to be
studied are stationary, are not ideal when applied to EEG data analysis. While traditional adaptive parameter estimation algorithms for
example the recursive least squares (RLS), least mean squares (LMS)
and Kalman ler approaches can be applied to track time-varying
trends, these algorithms can often produce lagged tracking of timevarying parameters (for example in TV-ARX model estimation). The
objective of this study is thus twofold: (i) to develop a time-varying
modelling framework that exploits the excellent approximation
property of wavelets so that the resultant time-varying models
can track severely nonstationary processes with sharp changes in
the system parameters; and (ii) to apply the time-varying modelling algorithm to EEG data analysis where the proposed algorithm
can provide almost instant tracking performance and the resultant timefrequency resolution produced by the identied model is
extremely high; this is usually very important for feature extraction
from EEG signals in both the time and frequency domains.
2. Problem formulation
2.1. Time-varying ARX model and multi-wavelet coefcient
expansions

p

i=1

ai (t)y(t i) +

q

n=1

bn (t)u(t n) + e(t),

(m)

k

hold:
ai (t) =

m = 3, 4, 5; k m , such that the following expression

()

()

i,k k

  
t

k 

bn (t) =

(r)

(r)

i,k k

  
t
N

k

() ()
n,k k

r
  
t

k 

(s)

(s)

 

i,k k

k

(r) (r)
n,k k

s
  
t

k r

(s)

(s)

n,k k

t
N

 
t
N

(2)

k s

where i,k and n,k represent the expansion parameters,

 m ={k : m k 2j 1} for m = 3, 4, 5, j = 3 is the wavelet scale,
 = 3, r = 4, and s = 5, t = 1, 2, . . ., N, and N is the number of observations of the measurement data, respectively. Substituting (2) into
(1), it yields,
y(t) =

p


i=1

q

n=1

(m)
i,k

(m)
k

t

k m

k m

(m)
n,k

(m)
k

y(t i)

t
N

u(t n)


,

(3)

From (3), the original TVARX model in Eq. (1) has now been converted into a time invariant (LTI) regression model with respect to
(m)
(m)
the time invariant coefcients i,k and n,k . In this study, cardinal B-splines wavelets, which have been proved to have several
excellent properties including orthogonality, support, regularity
and timefrequency localization in a window which enables the
operation of the multi-resolution decomposition to be much more
convenient, are considered and will be employed for time-varying
parameter expansion. Detailed discussions about how to build the
associated multi-wavelet model using B-splines can be found in
(Wei and Billings, 2002) and (Wei et al., 2010).
2.2. Time-dependent spectrum estimation
Eq. (3) can be solved by using linear least squares algorithms.
Let a i (t), b n (t) be the estimates of ai (t) and bn (t), and  e2 is the estimate of e2 . The time-dependent spectral function associated to the
TVARX model in Eq. (1) is dened as,

2
q

n (t)ej2nf/fs
b

n=1
H(f, t) =

,
1 pi=1 a i (t)ej2if/fs

(4)

where j = 1, and fs is the sampling frequency. Note that the

spectral function (4) is continuous with respect to the frequency f
and thus can be used to produce spectral estimates at any desired
frequency up to the Nyquist frequency fs /2. The frequency resolution is not innite, but is determined by the underlying model order
and the associated parameters.
2.3. Model identication and parameter estimation

The TVARX (p, q) model for a single-input/output system can be

represented as
y(t) =

where t is the time instant or sampling index of the signal y(t),

y(t i) and u(t n) are the measured response, respectively. ai (t)
and bn (t) are the TV coefcient functions to be determined in the
model; the term e(t) is the residual error accommodating the effects
of measurement noise, and modelling noise that can be viewed as
a stationary white noise sequence with zero mean and variance
e2 . The proposed method is to expand the TV parameters ai (t) and
b
n (t) onto multi-wavelet families
 cardinal B-splines basis functions,

(1)

In general, the estimation of LTI system parameters is formulated as an overdetermined problem. Then the least squares
solution is the optimal estimate of the parameters in the sense
of minimum residual error. However, if the parameters are
time-varying, the problem of parameter estimation becomes
underdetermined, and it is much more difcult to nd the best

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

153

Fig. 1. One implementation of the TVARX (2, 2) system identication results with a SNR of 13 dB using the different approaches. Blue curve represents the true value of the
TV parameters; red curve indicates the estimation value of the TV parameters. (a) The RLS method; (b) the RLS method with B-spline; (c) the OLS method with B-spline. (For
interpretation of the references to colour in this gure legend, the reader is referred to the web version of the article.)

solution. Expanding the TV parameters onto a linear combination of a set of basis functions can solve the underdetermined
problem. Consequently, the parameter estimation of unknown
variables can be reduced to a set of constant coefcients of the
basis functions. However, the multi-wavelet expansion model
(3) involves a large number of candidate model terms that may
be highly correlated. The resultant parameter estimates may
be over-tted. Experience suggests that most of the candidate
model terms can be removed from the model, and that only

a small number of signicant model terms are needed to provide a satisfactory representation for most linear and nonlinear
dynamical systems. Many approaches have been introduced to
eliminate the possible linear dependency of candidate model
terms by selecting best bases, for example, Kaipio and Karjalainen
(1997) introduced a principal-component-analysis (PCA)-type
approximation scheme to select the optimal basis. The mutual
correlation of the coefcients is also taken into account in their
approach.

154

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

Table 1
A comparison of the model performance for TVARX (2, 2) model with SNR 13 dB.
Approach

Estimated coefcient

MAE

RMSE

Std

RLS
( = 0.92)

a 1 (t)
a 2 (t)
b 1 (t)
b 2 (t)

0.0917
0.1030
0.1080
0.0627

2.3104
1.8667
1.2315
0.7260

0.1199
0.1292
0.2172
0.0863

RLS with
B-spline
( = 0.9998)

a 1 (t)
a 2 (t)
b 1 (t)
b 2 (t)

0.2045
0.2047
0.1411
0.1803

3.0639
1.9951
1.3973
1.0434

0.2623
0.2746
0.2084
0.2900

OLS with
B-spline

a 1 (t)
a 2 (t)
b 1 (t)
b 2 (t)

0.0893
0.0614
0.0642
0.0246

2.1750
1.3520
0.8638
0.2250

0.1112
0.0865
0.1133
0.0305

where e(t) is zero-mean Gaussian white noise. The TV parameters

in Eq. (5) are given by:

a1 (t) =

0.32 cos(1.5 cos(4t/N + )), 1 t N/4

0.32 cos(3 cos(4t/N + /2)),

N/4 + 1 t 3N/4 ,

0.32 cos(1.5 cos(4t/N + )),

3N/4 + 1 t N

a2 (t) = 0.4 cos(4t/N), 1 t N,

b1 (t) =

0.65, 1 t N/4

0.5,

(6)

N/4 + 1 t N/2

0.65, N/2 + 1 t 3N/4

0.5,

3N/4 + 1 t N

b2 (t) = 0.6, 1 t N,

where  represents the forgetting factor.

where the length of data N is 512. Model (5) was simulated by setting the input u(t) as a Pesudo-Random Binary Sequence (PRBS)
(Leontaritis and Billings, 1987). The variance of the noise e(t) was
chosen to be 0.04, and this made the signal-to-noise ratio to be
around 13 dB. Both the input and the associated output sequences
were recorded and were used for subsequent model estimation.
Fig. 1 compares three different methods, that is, the RLS algorithm, the RLS algorithm with B-spline basis functions, and the
OLS algorithm with B-spline basis functions. Panel (a) shows the
results using the RLS estimation algorithm (forgetting factor (ff)
0.92). Panel (b) gives the results of the RLS (ff: 0.998, using B-spline
wavelets and selecting scale index j = 3) algorithm and Panel (c)
shows the OLS identication results (using B-spline wavelets and
selecting scale index j = 3). Obviously, The RLS approach attains
smooth but relatively poor estimates that cannot track the rapidly
changing TV parameters, the parameter estimates are underdetermined. The RLS approach with B-spline obtains irregular estimates
with large variances (over-tted), however, compared with RLS
method, the resultant estimates from the RLS method with B-spline
can track the sharp changes of the TV parameters. These interest-

In this work, TV coefcients are expanded by multi-resolution

cardinal B-splines wavelet series, and then the forward orthogonal
least squares (OLS) algorithm (Chen et al., 1989; Billings et al., 1989,
2007; Wei and Billings, 2008), which have been proven to be a very
effective to deal with multiple dynamical regressions problems, is
applied to determine the forms in model (3). Detailed discussions of
the procedure of the forward OLS can be found in Chen et al. (1989),
Billings et al. (1989, 2007), Wei et al. (2004), Wei and Billings (2008).
The TV coefcients ai (t) and bn (t) in Eq. (1) can be recovered by the
resultant estimates from model (3).

3. Simulation example
Consider a TVARX (2, 2) model below
y(t) = a1 (t)y(t 1) + a2 (t)y(t 2) + b1 (t)u(t 1)
+ b2 (t)u(t 2) + e(t)

(5)

Input Signal

EEG(uV)

400
200
0
-200
-400

Output Signal

EEG(uV)

500

-500

Sample Time(sec)
Fig. 2. The EEG recordings (F3 Channel: input signal, F4 Channel: output signal), for a seizure activity of a patient, recorded over 7 s, with a sampling rate of 500 Hz.

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

155

Fig. 3. Estimates of the time-varying coefcients ai (t) for i = 1, 2, 3, 4 and bn (t) for the EEG signal.

ing results have been veried by Li et al. (in press). The OLS method
with B-spline appears to outperform the RLS approach and the RLS
approach with B-spline. The results using the OLS approach with Bspline is impressive because it is able to track three quite different
waveforms: the constant value, an abrupt change, and the sinusoidal waveform. The proposed method (the OLS with B-spline) can
attain smooth estimates while providing rapid tracking. The mean
absolute error (MAE), normalized root mean squared error (RMSE)
and the standard deviations (Std) of the parameter estimates

(with respect to the true parameters) are estimated and shown in

Table 1.
Compared with the RLS approach and the RLS approach with Bspline estimates, Table 1 statistically conrms that the MAE, RMSE
and Std estimates produced by the OLS approach with B-spline yield
smallest. The MAE and RMSE are both dened by

1 
a (t) a(t) ,
N
N

MAE =

(7)

t=1



 N 
 1  a (t) a(t)2

RMSE =
 2 ,
N
a(t)
t=1

Fig. 4. A comparison between the recovered signal from the identied TVARX (4, 3)
model and the original observations for the EEG signal. Solid (blue) line represents
the observations and the dashed (red) line represents the signal recovered from the
TVARX (4, 3) model. For a clear visualization only the data points of the period from 0
to 2 s are shown. (For interpretation of the references to colour in this gure legend,
the reader is referred to the web version of the article.)

(8)

Fig. 5. The 3-D topographical map of the time-dependent spectrum estimated from
the TVARX (4, 3) model for the EEG signal.

156

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

Fig. 6. The 2-D image and the contour diagram of the time-dependent spectrum produced by the 3-D topographical map. (a) The 2-D image; (b) the contour diagram.

where a (t) represents estimates of coefcients a(t) in the TVARX

model (3), and N is the length of the data set.
4. ApplicationEEG data modelling and analysis
4.1. Time-varying ARX model for EEG
The proposed TVARX modelling scheme has been applied to
analyse dynamic relationships from EEG recordings to illustrate the
application of the proposed multi-wavelet basis function method
based on TV parametric modelling. Scalp EEG signals are synchronous discharges from cerebral neurons detected by electrodes
attached to the scalp. The EEG signals considered here were
recorded with the same 32-channel amplier system. An XLTEK
32 channel headbox (Excel-Tech Ltd) with the international 1020
electrode placement system was used in the Shefeld Teaching
Hospitals NHS Foundation Trust, Royal Hallamshire Hospital, UK.
The sampling frequency of the device was 500 Hz. Andrzejak et al.
(2001) has discussed in detail dynamical properties of brain electrical activity from different extracranial and intracranial recording
regions and from different physiological and pathological brain
states. The central objective of this paper for the EEG signals is
to propose an empirical and data-based modelling framework

Fig. 7. The EEG recordings of C3 Channel, for a seizure activity of a epileptic patient,
recorded over 20 s, with a sampling rate of 500 Hz.

from model identication that can produce an accurate but simple description of the dynamical relationships between different
recording regions during brain activity. This is a complicated black
box system where the true model structure is unknown, and thus,
needs to be identied from available experimental data. As an
example, the symmetrical two channels (F3, located over the left
superior frontal area of the brain, and F4 located over the same area
on the right) of EEG recorded from a patient during an absence
seizure was investigated. Channel F3 was treated as the input,
denoted by u(t), and Channel F4 was treated as the output, denoted
by y(t), note that Channel F3 is the signal input and Channel F4
is the signal output, the main reason is that the phase of Channel
F4 is related to the phase of Channel F3. The objective is to investigate, from the available Channel F3 and Channel F4 recordings,
if an identied TVARX model is suitable to describe the dynamical characteristics by using the time-dependent spectrum analysis
approach. A data set, consisting of 3500 inputoutput data point
pairs of EEG signals representing one seizure epoch recorded for
a total of 7 s with a sampling rate of 500 Hz, was analysed in this
example. The plots of the associated samples are shown in Fig. 2.
Similar to the simulation example given in Section 3, the third,
fourth and fth order B-splines were adopted to establish TVARX
models for the EEG recordings. Several TVARX models with different model orders were estimated using the OLS approach with
B-spline, the classical generalized cross-validation (GCV) criteria
(Billings et al., 2007) suggested that the model order can be cho-

Fig. 8. The 3-D topographical map of the time-dependent spectrum estimated from
the TVAR (4) model for the EEG signal given in Fig. 7.

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

157

Fig. 9. The 2-D image and the contour diagram of the time-dependent spectrum produced by the 3-D topographical map for the EEG signal given in Fig. 7. (a) The 2-D image;
(b) the contour diagram.

sen to be p = 4 and q = 3 when using the B-splines as building

blocks to represent the time-varying coefcients in the TVARX
model.
The time-varying coefcients estimated ai (t) with i = 1, 2, . . ., 4
and bn (t) with n = 1, 2, 3 are depicted in Fig. 3. Fig. 4 shows the recovered signal, recovered by the TVARX model from the estimated
time-varying coefcients ai (t) and bn (t). The topographical diagram
of the time-dependent spectrum estimated from the TVARX (4, 3)
model is shown in Fig. 5, and the 2-D image diagram and the contour plot of the time-dependent spectrum produced from the 3-D
topographical diagram are given in Fig. 6.
From Figs. 5 and 6, the distribution scale of the power spectrum of the EEG signal considered here is mainly from zero to
around 18 Hz. Two frequency bands can clearly be observed as:
(a) the low frequency band (about 3 Hz); (b) around 18 Hz represents the high frequency band component. The contour plot of
the time-dependent spectrum given in Fig. 6(b) clearly reects the
distribution of these frequency components along with the time
course. It is clear that the variations of the time course signals can
be observed from the contour diagram of the transient spectrum.
For instance, the power spectrum is mainly distributed a 3-Hz frequency component during the period from 5 to 6 s, while the high
frequency (around 18 Hz) activity is dominated by the time course
from 0.2 to 0.3 s. Any time-invariant parametric modelling framework such as the commonly applied ARX models cannot attain
these properties which are only possessed by the TVARX model
proposed.
Note that in order to avoid any distortion of the EEG signals,
no ltering was applied to pre-process the raw EEG signals here
(and in the example below) apart from the notch frequency (50 Hz)
being removed.
4.2. Time-varying AR model for EEG
This example presents some illustrations to demonstrate the
applicability of the proposed modelling framework for characterising epileptic seizure EEG signals by using time-varying AR models.
Again, Channel C3 of the EEG signal of a subject for an epileptic
seizure activity lasting 20 s, recorded from a 10-year-old boy, was
considered. A total of 10,000 samples, recorded with a sampling
rate of 500 Hz, are shown in Fig. 7.
Similar to the previous example, the third, fourth and fth
order B-splines were adopted to establish TVAR models for the
EEG recordings. Numerical experiments have suggested that a
TVAR model of order p = 4 is appropriate for representing the EEG
recordings here. The topographical diagram of the time-dependent
spectrum estimated from the TVAR (4) model is shown in Fig. 8, and
the 2-D image diagram and the contour plot of the time-dependent

spectrum produced from the 3-D topographical diagram are given

in Fig. 9.
From Figs. 8 and 9, the power spectrum of the EEG signal with
an epileptic seizure activity of 20 s is mainly distributed in the frequency range from zero to 10 Hz. Three frequency bands, i.e. Delta
band (04 Hz), Theta band (47 Hz) and Alpha band (812 Hz), can
clearly be observed as follows: (a) the low frequency band (around
3 Hz); (b) the frequency band that is centralized around 6 Hz; (c)
around 10 Hz representing the high frequency band component.
The image and contour plot of the time-dependent spectrum shown
in Fig. 9 clearly reects the distribution of these frequency components along with the time course. It is clear that the variations of the
time course signals can be observed from the contour diagram of
the transient spectrum. For example, the power spectrum is mainly
dominated by a 3-Hz frequency component during the period of the
rst 2 s, 78 s, and 1112 s, respectively, while the high frequency
(around 10 Hz) activity is dominated by the time course from 16 to
18 s. Any time-invariant parametric modelling framework such as
the commonly applied conventional AR models cannot reveal these
properties.
5. Conclusions
A novel TV parametric modelling approach has been presented
based on time-dependent coefcients approximated multi-wavelet
basis functions to account for the transient spectrum information.
The TVARX model in this study is completely different from existing TV parametric models where the associated time-dependent
coefcients are expanded by multi-wavelet basis functions. In
most existing TV parametric models, the time-dependent coefcients represented by the basis functions are global, whereas in
the new proposed modelling method, the basis functions involved
are locally dened. Wavelets have been proved to show excellent
approximation properties (Mallat, 1989), therefore, the TV models
established by the multi-wavelet basis function expansion scheme
can be much more adaptable and exible for tracking the sharp
variations of nonstationary biomedical signals such as EEG recordings.
The time-dependent spectrum based on TVARX and TVAR
model, with multi-wavelet basis functions, can reect the global
frequency behaviour of the signal and to reveal the local variations
of the signal along the time course. One advantage of the proposed
model, compared with traditional time-invariant models, is that
it can capture much more transient information of the inherent
nonstationary dynamics of the associated processes.
Future studies will investigate how to exploit and integrate the
advantages of conventional adaptive estimation algorithms (e.g.
the least means squares and recursive least squares) and the excel-

158

Y. Li et al. / Journal of Neuroscience Methods 196 (2011) 151158

lent approximation property possessed by wavelets to produce

improved performance for EEG data analysis. It is believed that the
proposed time-varying modelling framework, coupled with results
obtained from some other useful nonparametric methods for example higher-order statistics approach (Zhou et al., 2008), can promise
signicant new results which can reveal new and important features buried in EEG signals, which can in turn enable further studies
and analysis, for example for disease pattern recognition and classication, detection and diagnosis. In addition, it is well known that
EEG signals are nonlinear processes, and thus linear models may not
be sufcient to represent such nonlinear processes; nonlinear representations should be more suitable for EEG data modelling and
analysis. Nonlinear modelling and analysis which has been under
our investigation will be discussed in a future separate publication.
Acknowledgements
Yang Li acknowledges the support provided by the University
of Shefeld under the scholarship scheme and the authors gratefully acknowledge that this work was supported by the Engineering
and Physical Sciences Research Council (EPSRC), UK, and European
Research Council (ERC).
References
Adeli H, Zhou Z, Dadmehr N. Analysis of EEG records in an epileptic patient using
wavelet transform. J Neurosci Methods 2003;123(1):6987.
Andrzejak RG, Lehnertz K, Mormann F, Rieke C, David P, Elger CE. Indications of
nonlinear deterministic and nite-dimensional structures in time series of brain
electrical activity: dependence on recording region and brain state. Phys Rev E
2001;64(6):061907.
Ansari AK, Bellanger JJ, Bartolomei F, Wendling F, Senhadji L. Timefrequency
characterization of interdependencies in nonstationary signals: application to
epileptic EEG. IEEE Trans Biomed Eng 2005;52(7):121826.
Billings SA, Chen S, Korenberg MJ. Identication of MIMO non-linear systems
using a forward-regression orthogonal estimator. Int J Control 1989;49(6):
215789.
Billings SA, Wei HL, Balikhin MA. Generalized multiscale radial basis function networks. Neural Networks 2007;20(10):108194.
Burke DP, Kelly SP, Chazal Pd, Reilly RB, Finucane C. A parametric feature extraction
and classication strategy for braincomputer interfacing. IEEE Trans Neural
Syst Rehab Eng 2005;13(1):127.
Cakrak F, loughlin PJ. Multiple window time-varying spectral analysis. IEEE Trans
Signal Process 2001;49(2):44853.
Chen S, Billings SA, Luo W. Orthogonal least squares methods and their application
to non-linear system identication. Int J Control 1989;50(5):187396.
Chen W, Chowdhury FN. Simultaneous identication of time-varying parameters
and estimation of system states using iterative learning observers. Int J Syst Sci
2007;38(1):3945.

Chon KH, Zhao H, Zou R, Ju K. Multiple time-varying dynamics analysis using

multiple sets of basis functions. IEEE Trans Biomed Eng 2005;52(5):956
60.
Chowdhury FN. Inputoutput modeling of nonlinear systems with time-varying
linear models. IEEE Trans Autom Control 2000;45(7):13558.
Chowdhury FN. A new approach to real-time training of dynamic neural networks.
Int J Adapt Control Signal Process 2003;17(6):50921.
Chui CK. An introduction to wavelets. Academic Press: London; 1992.
Jason DF, Bang VB, Algis JV. Wavelet analysis reveals dynamics of rat oscillatory
potentials. J Neurosci Methods 2008;169(1):191200.
Jiang B, Chowdhury FN. Fault estimation and accommodation for linear MIMO
discrete-time systems. IEEE Trans Control Syst Technol 2005;13(3):4939.
Kaipio JP, Karjalainen PA. Estimation of event-related synchronization changes by a
new TVAR method. IEEE Trans Biomed Eng 1997;44(8):64956.
Leontaritis I, Billings SA. Experimental-design and identiability for nonlinearsystems. Int J Syst Sci 1987;18(1):189202.
Li Y, Wei HL, Billings SA. Identication of time-varying systems using
multi-wavelet basis functions. IEEE Trans Control Syst Technol,
doi:10.1109/TCST.2010.2052257, in press.
Ljung L, Gunnarsson S. Adaptation and tracking in system identicationa survey.
Automatica 1990;26(1):721.
Mallat SG. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans Pattern Anal Mach Intell 1989;11(7):67493.
Nells O. Nonlinear system identication: from classical approach to neural networks
and fuzzy models. Springer-Verlag: Berlin, Heidelberg; 2001.
Ng SC, Raveendran P. Enhanced mu rhythm extraction using blind source separation and wavelet transform. IEEE Trans Biomed Eng 2009;56(8):2024
34.
Niedzwiecki M. Identication of time-varying process. New York: Wiley; 2000.
Pachori RB, Sircar P. EEG signal analysis using FB expansion and second-order linear
TVAR process. Signal Process 2008;88(2):41520.
Tsatsanis MK, Giannakis GB. Time-varying system identication and model validation using wavelets. IEEE Trans Signal Process 1993;41(12):351223.
Wei HL, Billings SA. Identication of time-varying systems using multiresolution
wavelet models. Int J Syst Sci 2002;33(15):121728.
Wei HL, Billings SA. Identication and reconstruction of chaotic systems using multiresolution wavelet decompositions. Int J Syst Sci 2004;35(9):51126.
Wei HL, Billings SA, Liu J. Term and variable selection for non-linear system identication. Int J Control 2004;77(1):86110.
Wei HL, Billings SA. An efcient nonlinear cardinal B-spline model for high
tide forecasts at the Venice lagoon. Nonlinear Process Geophys 2006;13(5):
57784.
Wei HL, Billings SA. Model structure selection using an integrated forward orthogonal search algorithm assisted by squared correlation and mutual information.
Int J Model Identication Control 2008;3(4):34156.
Wei HL, Zheng Y, Pan Y, Coca D, Li LM, Mayhew JEW, et al. Model estimation of
cerebral hemodynamics between blood ow and volume changes: a data-based
modelling approach. IEEE Trans Biomed Eng 2009;56(6):160116.
Wei HL, Billings SA, Liu J. Time-varying parametric modelling and time-dependent
spectral characterisation with applications to EEG signals using multi-wavelets.
Int J Model Identication Control 2010;9(3):21524.
Zhou SM, Gan JQ, Sepulveda F. Classifying mental tasks based on features of
higher-order statistics from EEG signals in braincomputer interface. Inf Sci
2008;178(6):162940.
Zou R, Chon KH. Robust algorithm for estimation of time-varying transfer functions.
IEEE Trans Biomed Eng 2004;51(2):21928.