IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO.

8, AUGUST 2015

1937

Matching Pursuit-Based Time-Variant Bispectral

Analysis and its Application to Biomedical Signals
Karin Schiecke , Member, IEEE, Matthias Wacker, Franz Benninger, Martha Feucht, Lutz Leistritz,
and Herbert Witte, Member, IEEE

AbstractObjective: Principle aim of this study is to investigate the performance of a matching pursuit (MP)-based bispectral
analysis in the detection and quantification of quadratic phase
couplings (QPC) in biomedical signals. Nonlinear approaches such
as time-variant bispectral analysis are able to provide information
about phase relations between oscillatory signal components. Methods: Time-variant QPC analysis is commonly performed using Gabor transform (GT) or Morlet wavelet transform (MWT), and is affected by either constant or frequency-dependent timefrequency
resolution (TFR). The matched Gabor transform (MGT), which
emerges from the incorporation of GT into MP, can overcome this
obstacle by providing a complex timefrequency plane with an individually tailored TFR for each transient oscillatory component.
QPC analysis was performed by MGT, and MWT was used as
the state-of-the-art method for comparison. Results: Results were
demonstrated using simulated data, which present the general case
of QPC, and biomedical benchmark data with a priori knowledge
about specific signal components. HRV of children during temporal
lobe epilepsy and EEG during burstinterburst pattern of neonates
during quiet sleep were used for the biomedical signal analysis to
investigate the two main areas of biomedical signal analysis: The
cardiovascularcardiorespiratory system and neurophysiological
brain activities, respectively. Simulations were able to show the applicability and reliability of the MGT for bispectral analysis. HRV
and EEG analysis demonstrate the general validity of the MGT
for QPC detection by quantifying statistically significant time patterns of QPC. Conclusion and Significance: Results confirm that
MGT-based bispectral analysis provides significant benefits for the
analysis of QPC in biomedical signals.

ECG
EEG
FFT
GT
HF
HR
HRV
LF
MAM
mBA
mBC
MGT
MP
MWT
QPC
QRS
QS
ROI
RSA
STFT
TFD
TFR
TLE
tvPS
WVD

Electrocardiogram.
Electroencephalogram.
Fast Fourier transform.
Gabor transform.
High frequency.
Heart rate.
Heart rate variability.
Low frequency.
Multiplicative amplitude modulation.
Mean biamplitude in the ROI.
Mean bicoherence in the ROI.
Matched Gabor transform.
Matching pursuit.
Continuous Morlet wavelet transform.
Quadratic phase coupling.
Q-, R-, and S-ECG-waves.
Quiet sleep.
Region of interest.
Respiratory sinus arrhythmia.
Short-term Fourier transform.
Timefrequency distribution.
Timefrequency resolution.
Temporal lobe epilepsy.
Time-variant power spectrum (spectrogram).
WignerVille distribution.

Index TermsElectroencephalogram, epilepsy, heart rate variability (HRV), quadratic phase coupling (QPC), time-variant bispectral analysis.

I. INTRODUCTION
AAM
AM
BA
BC

NOMENCLATURE
Additive amplitude modulation.
Amplitude modulation.
Biamplitude.
Bicoherence.

Manuscript received August 26, 2014; revised December 17, 2014 and February 6, 2015; accepted February 14, 2015. Date of publication February 26, 2015;
date of current version July 15, 2015. This work was supported by the DFG under
Grant Wi 1166/12-1 Le 2025/6-1. Asterisk indicates corresponding author.
K. Schiecke is with the Institute of Medical Statistics, Computer Sciences
and Documentation, Jena University Hospital, Friedrich Schiller University,
Jena 07740, Germany (e-mail: Karin.Schiecke@med.uni-jena.de).
M. Wacker, L. Leistritz and H. Witte are with the Institute of Medical Statistics, Computer Sciences and Documentation, Jena University Hospital, Friedrich
Schiller University Jena, Germany.
F. Benninger is with the Department of Child and Adolescent Neuropsychiatry, and M. Feucht is with the Department of Child and Adolescent Medicine,
University Hospital Vienna, Austria.
Preliminary results were published in a conference paper at the 36th Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, IL,
USA, August 2630, 2014.
Digital Object Identifier 10.1109/TBME.2015.2407573

UADRATIC phase coupling (QPC) is an important property of biomedical signals indicating a specific nonlinear
coupling between two oscillatory signal components. For example, if a signal with only two sinusoidal (oscillatory) components
(frequencies f1 and f2 with f2 > f1 ) passes through a system
with a quadratic nonlinearity [see (15)], then the output signal contains frequency components at f1 , f2 , 2f1 , 2f2 , f2 f1 ,
and (zero) phase relations are of the same type as the frequency relations, i.e., 1 , 2 , 21 , 22 , 2 1 [1]. Such a
coupling configuration is called QPC. Higher order spectral
analysis is capable of detecting and characterizing QPC [2]. A
frequency triplet {f1 , f2 , f3 } produces a high peak in the bispectrum B (fm , fn ) (bispectral plane) at the coordinates [f1 , f2 ]
if the coupling conditions f3 = f2 + f1 and 3 = 2 + 1 are
fulfilled, i.e., this coincides with the frequency and phase conditions defined as QPC. Additionally, bispectral peaks at the diagonal of the bispectral plane (coupling of the first and second harmonics) and at the coordinates [f1 , f2 f1 ] occur because for
the component f2 f1 , the triplet-related coupling conditions

0018-9294 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

1938

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

f3 = (f2 f1 ) + f1 = f2 and 3 = (2 1 ) + 1 = 2
are also met. For f2  f1 , the bispectral peaks at [f1 , f2 ] and
[f1 , f2 f1 ] are in a tight neighborhood, i.e., they can be only
distinguished by using a high-frequency (HF) resolution.
Stationary (time invariant) bispectral analysis has already
been implemented for many years [3]. Biomedical signals are
usually nonstationary, i.e., QPC in biomedical signals change
in time and a high time resolution is desirable. This is the reason why time-variant implementations of bispectral analysis
are standard for biomedical applications. The categorization of
time-invariant bispectral techniques (stationarity of the signal is
required) can also be used for time-variant versions. One can distinguish between nonparametric (conventional [1]) and parametric approaches. Frequently used nonparametric approaches
are based on the short-term Fourier transform (STFT) [4], Gabor transform (GT) [5], Morlet wavelet transform (MWT) [6],
on third-order timefrequency distribution (TFD) (Wigner bispectrum [7] and its advanced versions [8], [9]). Time-variant
parametric bispectral approaches require an estimation of parameters for a time-variant autoregressive (AR) model, which
consider higher order moments (e.g., [10]). For bispectral analysis of interval signals (derived from point processes), the nonlinear AR integrative model and other nonlinear AR approaches
have been introduced [11], [12].
A crucial point for the application of all time-variant bispectral techniques is their timefrequency resolution (TFR). It
would be beneficial if for each signal component an appropriate (individually adapted) TFR would be available in order to
separate existing components from each other (frequency resolution) and to localize QPC changes in time (time resolution) [6].
However, localization in time and localization in frequency are
contradictory aims (uncertainty principle), i.e., the decrease of
time resolution results in an increase of the frequency resolution
and vice versa (for an overview, see [13]), i.e., when interpreting results of time-variant bispectral analysis, it is mandatory to
understand the specific effects of TFR. The TFR properties of
the time-variant approaches mentioned above are summarized
in Section V.
Summarizing these facts, it can be assumed that the availability of individually tailored TFRs, i.e., TFRs, which are individually adapted to the properties of each signal component
(signal adaptive), would be a big advantage for bispectral analysis in order to detect and quantify the temporal organization of
QPCs. The matched Gabor transform (MGT) [14] was recently
introduced to obtain a linear phase analysis with these beneficial
properties. MGT emerges from the incorporation of the GT into
the matching pursuit (MP) decomposition algorithm. For the
MGT, the MP with a dictionary of real-valued Gabor atoms is
used (cosine multiplied with a Gauss function). In general, the
MP [15] decomposes a signal into a sum of atoms from a given
dictionary. The difference between the original MP (MP plus
WignerVille distributionWVD) and the MGT lies in generation of the aggregated timefrequency plane. Every selected
atom is used to define an individual time window (time resolution), and thereby, the corresponding frequency resolution for
the GT is defined, i.e., individually tailored TFR results. The
resulting complex timefrequency planes with different TFRs

Fig. 1. General design of processing scheme. Frequency-dependent MWT as

well as MGT-based strands of analysis are depicted.

are then summed up to obtain the aggregated (complex) time

frequency plane. Amplitude and information are available for
each point of the aggregated (complex) timefrequency plane.
It should be mentioned here that WVD does not provide instantaneous phase information.
The main objective of this study can be summarized as follows: The application of the MGT leads to improved analysis
results for linear-phase properties [14]. An improvement can
also be expected for time-variant bispectral analysis but any
clear evidence is still pending. Consequently, we want to show
that MGT-based bispectral analysis has key advantages, which
can be beneficially used for QPC analysis. Therefore, the first
aim of this study is to provide the evidence of this by using
simulations, which are related to the QPC coupling conditions
which change abruptly. These simulations demonstrate both the
advantages of the approach yet also the restrictions and pitfalls. The results of the MGT-based QPC analysis are compared
to the MWT-based results. Second, the aim is to demonstrate
that this new approach yields improved results for bispectral
analysis of both heart rate variability (HRV) and EEG. By analyzing HRV as well as EEG data, two main application areas
of biomedical signal analysis are considered: The analysis of
the cardiovascularcardiorespiratory system and of neurophysiological brain activities. We have used data for which the timevariant QPC analysis on the basis of the MWT (state-of-the-art)
has been already performed, and therefore, a priori knowledge
about signal properties exists (benchmark data).
II. METHODS
The general design of the processing scheme is depicted
in Fig. 1. Investigated data, applied approaches, and their
main characteristics are given. Methodological backgrounds
of MWT, MGT, and time-variant bispectral approaches, as
well as descriptions of TFR and applied statistical analyses

SCHIECKE et al.: MATCHING PURSUIT-BASED TIME-VARIANT BISPECTRAL ANALYSIS AND ITS APPLICATION TO BIOMEDICAL SIGNALS

1939

are given in Sections II-A to II-E. In general, the timevariant definition of QPC implies that a pair of two signal
components {x1 , x2 }, with x1 (t) = a1 sin(2f1 t + 1 ) and
x2 (t) = a2 sin(2f2 t + 2 ), is quadratic phase coupled when
a third component x3 (t) = a3 sin(2f3 t + 3 ) with the frequency f3 = f1 + f2 and the zero phase 3 = 1 + 2 exists. As the instantaneous phase of x3 (t) is 3 (t) = 2f3 t+
3 = 2f1 t + 2f2 t + 1 + 2 = 1 (t) + 2 (t), the QPCs
phase condition can be generalized by using instantaneous
phases instead of zero phases. The instantaneous phase triplet
{1 (t), 2 (t), 3 (t)} can be computed for each time point using
of timefrequency methods like GT and MWT.

Here, g is a complex-valued Gabor atom of the GT with the

general form


2 
t
1
g (t, f, , ) = exp (i (2f t + )) exp 0.5

2
(7)
k
and for
j

m in , jk m in

k = k ,
m in < jk < m ax
(8)

j
j

m ax , jk m ax

A. Morlet Wavelet Transform

with

The frequency-dependent complex analytic signal y k (t, fn )

of the data xk (t, f ) is computed using MWT [13], where k
designates the number of signals (realizations, trials). The timevariant power spectrum (tvPS) PSk (t, fn ) and phase k (t, fn )
can be calculated on the basis of the complex analytic signal for
each recording k by

2
(1)
PSk (t, fn ) = y k (t, fn )

b ounds = [m in , m ax ]

(9)

applies.
Regarding y k (t, fn ) in (5), the tvPS and phase k (t, fn )
can be calculated on the basis of this complex timefrequency
plane for each recording k, and again, the representative tvPS
by ensemble averaging according to (3). A detailed description
of MGT can be found in [14]
C. Time-Variant Bispectral Analysis

and
k (t, fn ) = arg y k (t, fn ) .

(2)

By ensemble averaging, the representative tvPS can be estimated

tvPS (t, fn ) =

K
2
1   k
y (t, fn ) .
K

(3)

k =1

The tvPS quantifies the averaged spectral characteristics of

our K realizations of QPC in a timefrequency map representation.
B. Matched Gabor Transform
MGT generates in a first step (MP) for each k, a linear combination of real-valued Gabor atoms dkj (t) (cosine term) from a
redundant dictionary D that approximates xk (t)
xk (t)

M


akj dkj (t) .

(4)

j =1

Second, each atom is analyzed with its own analysis time

k , to generate a set of timefrequency
window, determined by
j
planes yjk (t, fn ), and finally, theses planes are combined to obtain a timefrequency plane y k (t, fn )
y k (t, fn ) =

M


akj yjk (t, fn )

(5)



k , 0 d.
dkj ( ) g t , fn ,
j

(6)

j =1

where


yjk

(t, fn ) =

Time-variant QPC between frequency bands can be computed using time-variant biamplitude and bicoherence. According to [1], for each realization/seizure (k = 1, . . . , K), the following triple product can be calculated for every frequency pair
(fm , fn ) and at each point in time
B k (t, fm , fn ) = y k (t, fm ) y k (t, fn ) y k (t, fm + fn ).
(10)
Here, denotes the complex conjugate. The ensemble averaging of B (k ) (t, fm , fn ) yields an estimation of the time-variant
(t, fm , fn ).
bispectrum B
The time-variant bispectrum depends on the amplitudes of the
frequency components {fm , fn , fn + fm }. For detecting phase
couplings, the estimation of the time-variant bicoherence is used
as an amplitude-independent measure
(t, fm , fn ) =





B (t, fm , fn )

. (11)
 k
 k
2
1
k (t, f )|2
|y
(t,
f
)

y
|y
(t,
f
+
f
)|
m
n
m
n
K

To reduce the three dimensions of the time-variant bispectral measures resulting from (8) and (9), the mean biamplitude
(mBA) and the mean bicoherence (mBC) were computed in the
regions of interest (ROI) F1 F2 according to
n u1 n u2
(t, fi , fj )
B
i=n l1
j =n l2
(t) = 



mB
nu1 nl1 + 1 nu2 nl2 + 1
n u1 n u2
(t, fi , fj )
i=n l1
j =n l2
(t) =
m
(12)
u
l
(n1 n1 + 1) (nu2 nl2 + 1)
where nl1 , nu1 , nl2 , and nu2 denote the according indices of

1940

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

the frequency grid of the ROI of the estimated bispectrum (lower/upper bounds of F1 and F2 ). According to the
two different frequency ranges of phase coupling in our
simulated data, ROI1 was set to F1 = [0.025 Hz, 0.075 Hz]
and F2 = [0.15 Hz, 0.25 Hz], and ROI2 was set to F1 =
[0.075 Hz, 0.125 Hz] and F2 = [0.3 Hz, 0.4 Hz], respectively.
Regarding the main rhythms of our data, we calculated timevariant bispectral measures in the ROI. For HRV analysis, we investigated QPC between the Mayer wave-related low-frequency
(LF) component and the respiration-related HF component in
the HRV, i.e., the ROI was set to F1 = [0.075 Hz, 0.15 Hz]
and F2 = [0.25 Hz, 0.35 Hz]. For EEG analysis, we investigated QPC between the delta waves related component and
the theta wave related component of the burst pattern during
quiet sleep (QS), i.e., the ROI was set to F1 = [0.5 Hz, 1.5 Hz]
and F2 = [3 Hz, 4 Hz]. A more detailed description of parameter extraction in case of time-variant bispectral analysis can be
found in [10].
D. TimeFrequency Resolution
The most important difference between MWT and MGT consists in the resulting TFR: TFR of MWT is frequency dependent,
and TRF of MGT is signal adaptive.
For MWT, higher frequencies lead to a better time resolution t but also a disadvantaged frequency resolution f and
vice versa. The standard deviation of the Gauss envelope in the
time domain and the standard deviation of the Gauss curve in
the frequency domain can be used as a measure for the time
and frequency resolution [13]. For practical purpose, a freely
selectable parameter o has to be set, t (f ) is bound to the
o
frequency and o (t = 2
f = 1/f ). Parameter o is dependent on the choice of the Morlet mother wavelet. In a first step,
the mother wavelet of the MWT was adapted so that the sigma
parameter of its Gaussian envelope equals one cycle for every
frequency (o = 2; denoted by A in Section IV). The resulting TFR is adequate to analyze the time pattern of linear activity
but inadequate for bispectral analysis (frequency resolution is
too low; missing side bands; for reference, see [13]). Therefore, the mother wavelet has to be readapted to the situation. For
simulations and HRV data, we used o = 2 (denoted by A
in Figs. 4 and 6) and o = 10 (denoted by B in Figs. 4 and
6). For EEG data, we used o = 2 (denoted by A in Figs. 7
and 8) and o = 5 (denoted by B in Figs. 7 and 8).
For the case of MGT, the resulting timefrequency planes
have multiple intrinsic TFR and cannot be obtained by a single GT. Each signal part is analyzed with its matched analysis
function. In case of GT, t (t = 1/f ) is freely selectable.
MGT can be applied by using the so-called sigma bounds
(see b ounds in (9)). Gabor atoms of the dictionary with sigma
values below the lower bound (with time resolutions, which are
too high) will be computed (GT) with a fixed (bound value),
lower time resolution. Respectively, the same applies for Gabor atoms of the dictionary with sigma values above the upper
bound (time resolutions, which are too low: computation with a
fixed, higher time resolution). The use of a lower sigma bound
ultimately leads to a smoothing in time of results, which seems

to be appropriate for our investigation. Therefore, we used only

lower sigma bounds for our investigation of the simulations and
HRV data: b ounds = [5 s, Inf] (denoted by C in Figs. 4 and
6) and b ounds = [10 s, Inf] (denoted by D in Figs. 4 and
6), and for EEG data, respectively: b ounds = [0.1 s, Inf] and
b ounds = [0.25 s, Inf] (denoted by C and D in Fig. 7.
E. Statistical Analysis
The statistical hypothesis testing for time-variant mBC analysis of real data (HRV and EEG) was performed using two
different methods. First, a surrogate data approach was applied,
and second, a bootstrapping approach was utilized.
The null hypothesis of the surrogate data approach is that
there is no QPC between the two frequency components of the
ROI (small values of mBC). The surrogate data were obtained
by destroying the phase information of investigated signals by
means of phase randomization [16]. This was carried out for
1000 repetitions, and the 95th percentile of the surrogate mBC
in the ROI was computed. The mean over time 95th percentile
was set as the 5% threshold for statistically significant mBC
values (see e.g., [17]).
Aim of the bootstrapping approach is to determine statistical properties of the single mBC time courses estimated for K
realizations of real data according to (11) and (12). In order to
estimate confidence tubes of the extracted mBC time courses
without any particular distribution assumption 1000 bootstrap
samples of size K (each sample element contains 10-min epochs
for HRV data and 10-s epochs for EEG data, respectively) were
drawn by a case resampling with replacement. With it, 1000
bootstrap replications of each extracted parameters were computed. Based on these replications, the lower limit (the 2.5%
quantile) defined the lower bound, and the upper limit (the
97.5% quantile) defined the upper bound of the 95% confidence
tube of mBC. For theoretical details concerning bootstrapping
approaches, see [18].
III. SIMULATION AND DATA MATERIAL
A. QPC Simulations
The first simulation (SIMQPC ) was used to test the properties
of the MGT-based time-variant bispectral analysis for the case
of abrupt changes in QPC simulated by a superimposition of
sinusoidal signal components. The analysis of SIMQPC serves
to clarify the following three methodological issues:
1) Time dynamics of bispectral parameters can be analyzed.
2) The influence of different parameter settings can be assessed. The results obtained by using our new approach
are compared with those from MWT (state-of-the-art).
Two different parameter settings for each approach are
used in order to demonstrate the influence of different
mother wavelets (for MWT) and of different time resolution bounds (for MGT).
3) Criteria (rules) for the choice of ROIs, in which the parameters are determined, can be identified.
SIMQPC consists of K realizations of QPC. Structure, time
length, and sampling frequency of the simulations were adapted

SCHIECKE et al.: MATCHING PURSUIT-BASED TIME-VARIANT BISPECTRAL ANALYSIS AND ITS APPLICATION TO BIOMEDICAL SIGNALS

Fig. 2. Schematic view of frequency distributions and time courses of simulations. One realization of SIM Q P C is depicted in (a). In the upper row, schematic
amplitude spectra are given for the two different situations of phase coupling in
SIM Q P C . Arrows illustrate the actual phase-coupled frequency triplets. In the
lower row, the time course of change between both situations of SIM Q P C (four
periods of 150-s length, each divided by dashed black lines) is depicted. (b)
time course of SIM A M is given. A division between no phase coupling/phase
coupling (each period 30 s) is given by dashed black lines.

to real data (K = 18; sampling frequency: 2 Hz). Schematic

representation of frequency distribution and time course of
SIMQPC is given in Fig. 2(a). Each realization of simulated
signal y(t) is made of a sequence of two alternating 150-s segments, each of them containing one phase-coupled frequency
triplet and two uncoupled frequencies
 1
y (t) : 0150 s and 300450 s
(13)
y (t) =
y 2 (t) : 150300 s and 450600 s
with
j

y (t) =

5
i=1

Phases are chosen randomly for each realization and ai,j = 1

applies for i = 1, . . . , 5 and j = 1, 2.
Frequency values chosen for SIMQPC are characteristic of
the LF and HF HRV ranges which are defined as follows (Task
Force HRV [19]): LF = 0.04 0.15 Hz; HF = 0.15 0.4
Hz. Component 0.05 Hz lies at the lower end of the LF range
and 0.2 Hz at the lower end of the HF range; 0.1 Hz is characteristic of HRV waves, which are strongly associated with
the Mayer waves of the systemic blood pressure and 0.35 Hz
can be assumed to be the frequency of the respiratory sinus
arrhythmia (RSA) in children during rest. The QPCs between
0.05 and 0.2 Hz as well as between 0.1 and 0.35 Hz represent
two couplings between the LF and the HF range, where the last
is specifically related to a coupling between HRV-related Mayer
waves and RSA.
Second simulations were performed to evaluate the QPC
detection features of MGT in the case of fast changes between phase couplings/no phase couplings under the condition that QPC is generated by nonlinear models. QPC was
generated by using both additive (asymmetric) and multiplicative (symmetric) amplitude modulation (AM). AM phenomena have been already described in HRV and EEG (e.g., [20]
and [21]). Two sinusoidal signals x1 (t) = a1 sin(2f1 t + 1 )
and x2 (t) = a2 sin(2f2 t + 2 ) with a2 = 5 a1 , f1 = 1 Hz,
f2 = 3.5 Hz, and 1 = 0, 2 = /4 were used. A frequency
ratio between 1:3 and 1:4 roughly represents our previously
validated bounds for the ROI for QPC detection in both applications (HRV: [0.075 Hz, 0.15 Hz] [0.25 Hz, 0.35 Hz) [22];
EEG: [0.5 Hz, 1.5 Hz] [3, 5 Hz] [23]).
For additive amplitude modulation (AAM), both signals were
added and applied as input signal x(t) = x1 (t) + x2 (t) to a
system with a quadratic nonlinearity
yAAM (t) = x (t) + x2 (t)

(14)

For j = 1, phase-coupled triplet is {f1,1 = 0.05 Hz, f2,1 =

0.2 Hz, f3,1 = 0.35 Hz}, and uncoupled frequencies are set
to {f4,1 = 0.1 Hz, f5,1 = 0.35 Hz}. For j = 2, phase-coupled
triplet is {f1,2 = 0.1 Hz, f2,2 = 0.35 Hz, f3,2 = 0.45 Hz},
and uncoupled frequencies are set to {f4,2 = 0.05 Hz,
f5,2 = 0.2 Hz}, respectively. Phase relations are of the same
type as frequency relations (3,j = 1,j + 2,j for j = 1, 2).

(15)

where yAAM (t) is the resulting output signal ( = 0.5). The

simulated signal (hereinafter referred to as SIMAM ; Fig. 2(b))
was composed by alternating segments of yAAM (t) (30-s segment with QPC) and x(t) (30-s segment without QPC). AAM
corresponds exactly to the general QPC definition [1] because
yAAM (t) contains frequency components f1 , f2 , 2f1 , 2f2 , f2
f1 , and the instantaneous phase relations are of the same type
(1 (t), 2 (t), 21 (t) , 22 (t) , 2 (t) 1 (t)).
In case of a multiplicative amplitude modulation (MAM), the
resulting signal can be described by
yM AM (t) = [1 + c1 x1 (t)] x2 (t)

ai,j cos (2fi,j t + i,j ) .

1941

(16)

and the condition for the product of the constants c1 a1 < 1

must be fulfilled. The signal yM AM (t) contains components
at frequencies f2 , f2 f1 with the corresponding phase relations 2 (t), 2 (t) 1 (t). Alternating segments of yM AM (t) +
x1 (t) (30-s segment with QPC) and 0-s segment without QPC)
were used as the signal to be analyzed. Results from this specific
simulation study are not presented here, however are discussed
in conjunction with the AAM results.

1942

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

B. Real Data
1) HRV of Children During TLE: Data are derived from a
group of 18 children; each with one seizure recording of at least
10 min (K = 18 seizures; median age 9 years 4 months, interquartile range 8 years 7 months to 12 years 1 months, range
6 years 6 months to 18 years 0 months; median seizure length
88 s, interquartile range 72110 s, range 52177 s). Presurgical evaluation was performed at the Vienna pediatric epilepsy
center following a standard protocol. EEG was recorded referentially from gold-disc electrodes placed according to the extended
1020 system with additional temporal electrodes. One-channel
ECG was recorded from an electrode placed under the left clavicle. EEG and ECG data were recorded referentially against
CPZ , filtered (170 Hz), converted from analog to digital (sampling frequency 256 Hz), and stored digitally for further data
analysis. Video recordings of each seizure were reviewed to
classify seizure type. Complex partial seizures were included,
but not auras or generalized tonic-clonic seizures. Seizure onset
and termination were determined by the EEG (independently
by two neurologists experienced in the field of epilepsy and
clinical electrophysiology). EEG and ECG recordings including 10-min epochs [5 min before (preictal state) and 5 min after
seizure onset (seizure and postictal state)] were stored for each
seizure.
QRS detection was performed after bandpass filtering (10
50 Hz) and interpolation by cubic splines (interpolated sampling frequency 1024 Hz, according to [24]) to detect the time
point of the maximum amplitude of each R-wave and the resulting series of events was used for the HRV computation. The
low-pass-filtered event series was computed by applying the
FrenchHolden algorithm [25]. The final HRV representation
was obtained from the low-pass-filtered event series via multiplication with the sampling rate and with 60 beats/min and down
sampled to 2 Hz. An artifact rejection was performed manually
to minimize the influence of false QRS triggering.
All results were determined by grand mean analysis over 18
seizures (K = 18 children). A detailed investigation of the HRV
by means of time-variant and frequency-selective linear and
nonlinear methods of the same group of children was recently
published in this journal [22]. Original HR courses of all children
are depicted in Fig. 3(a).
2) EEG of Neonates With BurstInterburst Pattern During
QS: A group of six full-term neonates (mean conceptual age
39.3 weeks, range 3841 weeks; mean birth weight 3152 g,
range 26703420 g; mean 5 min APGAR-score 9, range 810)
was analyzed. Recordings were performed during sleep between
09.00 and 12.00 h; all neonates lay in an incubator at temperatures adapted to maintain normal body temperature and none
showed any EEG abnormality. Eight channel EEG (128-Hz sampling rate, international 1020 system with electrodes F p1 , F p2 ,
C3 , C4 , T3 , T4 , O1 , O2 ), heart rate, respiratory movements, and
EOG were recorded.
Only the EEG recorded during QS was selected. The EEG
was segmented by a trained physician, the burst onset was used
as a fix point for a 10-s interval. 4 s before (interburst) and 6 s
after the burst-onset were considered. For the visual detection
of the burst onset (amplitude criteria), the burst was defined as

Fig. 3. Examples of real data used for analyses. (a) HR of all children (K =
18; thin gray lines; upper row) and median HR (bold black line) with interquartile
tube (25% to 75% quartiles; gray filled) for all children (lower row) is depicted.
Distinct time points of TLE seizures are given: dashed line designates the onset
of the preonset acceleration (240 s), full line designates the EEG seizure onset
(300 s), dotted line designates the maximum of the acceleration and beginning
of deceleration (340 s), and dashed-dotted line designates the median end of
seizure (390 s). (b) All investigated EEG pattern (K = 17, thin gray line) and
first EEG pattern (bold black line) of one exemplary child (#1) are shown. Burst
onset is at 4 s (bold black line).

the simultaneous appearance of a group of high amplitude (>

50 V), LF waves (0, 53 Hz) in more than 75% of the recording channels. These waves are superimposed by low amplitude
(<50 V), HF waves (415 Hz). From our previous study [23],
10-s intervals of the F p1 recordings were analyzed for each
neonate; the burstinterburst patterns starting with the beginning of the QS period were selected for analysis (K = 17 for
mean analysis per neonate: minimal number of available 10-s
intervals in one neonate; K = 102 for grand mean analysis of
all neonates). Examples of the EEG pattern of one neonate are
depicted in Fig. 3(b).
IV. RESULTS
A. Simulations
A comparison of grand mean results of SIMQPC (K = 18
trials of simulations) of frequency-related tvPS between the
MWT and MGT is depicted in Fig. 4. The TFR from the MWTbased approach is dependent on frequency that is clearly visible
[see Fig. 4(a-A) and (a-B)]. It is not possible to correctly detect
the QPC in (A) because of the occurrence of cross terms between
higher frequencies (LF resolution). The TFR from all the MGT
approaches [see Fig. 4(a-C), and (a-D)] is adequate for bispectral
analysis. Representations of BA at two defined single time points

SCHIECKE et al.: MATCHING PURSUIT-BASED TIME-VARIANT BISPECTRAL ANALYSIS AND ITS APPLICATION TO BIOMEDICAL SIGNALS

1943

Fig. 4. Grand mean results (K = 18) of SIM Q P C . (a) tvPS, (b1) and (b2) BA at two different time points (75 and 225 s), and (c) mBA in the ROIs are shown. (A)
and (B) depict results of MWT-based approaches ( 0 = 2/10), and (C) and (D) depict results of MGT-based approaches ( b o u n d s = [5 s, Inf]/[10 s, Inf]).
Red color designates high power, blue color: low power of tvPS or BA, respectively, in column (a), (b1), and (b2). In column (c), light gray line depicts
time course of mBA in ROI1 = [0.025 Hz, 0.075 Hz] [0.15 Hz, 0.25 Hz], dark gray line in ROI2 = [0.075 Hz, 0.125 Hz] [0.3 Hz, 0.4 Hz]. ROIs are
framed in the adequate color in (b1) and (b2). For (D), the use of a wider ROI (guideline-driven according to [19]) for mBA calculation is depicted (ROI =
[0.025 Hz, 0.125] [0.15 Hz, 0.4 Hz]; dashed red rectangle in (b1, (b2), and red line in (c)).

(t1 = 75 s and t2 = 225 s; b1 and b2; see Fig. 1(a) for simulated
QPC at these time points) as well as the time course of mBA
in both investigated ROIs (c) are given in Fig. 4(b) and (c) for
simulated data and all investigated MWT [see (A) and (B)] and
MGT-based approaches [see (C) and (D)]. According to our
simulations of phase couplings, there should be a peak in the
biamplitude of f1 = 0.05 Hz and f2 = 0.2 Hz at the time point
presented in (b1) and a peak in the biamplitude of f1 = 0.1 Hz
and f2 = 0.35 Hz at the time point presented in (b2).
In general, all approaches were able to depict these peaks,
but there is an increasing frequency resolution (and, therefore,
decreasing time resolution) from (A) to (B) and (C) to (D).
In addition, due to the occurrence of couplings at lower frequencies, (A) to (B) for MWT is characterized by a higher
frequency resolution (and consequently lower time resolution)
at time point (b1) in comparison to (b2). Quantification of mBA
in the ROIs [see Fig. 4(c)] shows a correct time course for all
approaches, but a partially false quantification of strength of the
QPC [see (A) to (B)]. Note that the simulations show a defined
(and equal) strength of QPC. The simulations showed better results for the MWT by applying maximum and not mean BA in
the ROI. However, this procedure is not possible to use on real
data (sensitivity to outliers of BA/BC).

In Fig. 5, results of AAM-based simulation (SIMAM ) are

given. The fast changes (30 s, 3 cycles of frequency component
f1 ) in the occurrence of phase coupling/no phase coupling were
in the focus of interest in the analysis. To highlight both the
advantages and disadvantages of MGT, no bounds of jk [see
Fig. 5 (a)] and adapted b ounds [see Fig. 5(b) and (c)] were used.
The results of timefrequency related PS show that the MP is
able to decompose the AAM related segments into stationary
components of an AM (f1 , f2 , f2 f1 ). This is also true for the
MAM-based simulations (not shown).
The quality of the decomposition depends on the localization
of the AAM segment within the analysis interval: Quality for
segments at the beginning and at the end of the analysis interval
is somewhat low [see Fig. 5(a)(c)]. In contrast, for segments
located in the central area of the interval, all components can
be clearly detected. Without smoothing [see Fig. 5(a)], the peripheral segments are decomposed into short Gabor atoms that
correspond to the subsequent spindles of the AAM signal, i.e.,
they are artificially fractionized into periodically repeating
(f1 ) Gabor atoms (frequency f2 ).
Consequently, we used for our real-world applications an
MP approach with a slight smoothing (b ounds ), and we chose
segments of our data in such a way that the pattern changes

1944

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

Fig. 5. Result of MGT-based frequency related tvPS of SIM A M . (a) MGT

approach without bounds of jk is depicted, (b) b o u n d s = [5 s, Inf], and (c)
b o u n d s = [10 s, Inf]) are used, respectively.

are located in the midarea of the analysis interval, i.e., expected

QPC changes before and after the state/pattern change can be
reliably quantified.
B. Real Data
1) HRV of Children During TLE: A comparison of grand
mean results of all children (K = 18 recordings of the HRV)
from the MWT and MGT-based tvPS of HRV is depicted in
Fig. 6(a). Again, the frequency-dependence of TFR from the
MWT is clearly visible [see Fig. 6(a-A) and (a-B)]. Also, the
TFR of (A) from the real data is not able to depict the side bands
of frequencies of QPC necessary for successful bispectral analysis (missing side band at 0.3 0.1Hz: modulation of RSA
by Mayer waves), whereas the TFR from all of the MGT-based
approaches [see Fig. 6(a-C) and (a-D)] is adequate for bispectral analysis. Representation of BA at three defined time points
(240, 340, and 390 s; b1 to b3; see Fig. 3(a) for relevance of
distinct time points) as well as the time course of mBC in the
ROI (c) are given in columns (b1)(b3) and (c) of Fig. 6 for
HRV of children with TLE and all investigated MWT [see (A)
and (B)] and MGT-based approaches [see (C) and (D)]. Note
that for real data, the estimation of the amplitude-independent
mBC is necessary to quantify the occurrence of QPC in the time
course. At our distinct time points, there should be a peak in the
biamplitude in the ROI = [0.075Hz, 0.15Hz] [2.5Hz, 3.5Hz]
at 240 s [start of preonset acceleration; Fig. 6(b1)], no peak at
all at 340 s [maximum of acceleration; ictal period; Fig. 6(b2)],
and again an increasing BA at 390 s [end of seizure; Fig. 6(b3)].

Red lines in Fig. 6(c) quantify statistical significance of peaks

in the mBC in their time course.
In general, all approaches are able to depict peaks of BA,
but there is a strongly increasing frequency resolution (and,
therefore, decreasing time resolution) from A to B and not such
a strong increase from C to D. The frequency-dependence of
TFR in the MWT results in minimally interpretable BA maps
[see Fig. 6(A) and (B)]. The high frequency at 0.3 Hz (our
focus) has inadequat frequency resolution [see Fig. 6(A)] and
as the frequency resolution increases [see Fig. 6(B)], the time
resolution is reduced (time smooting). MGT-based BA maps are
easier to interpret, and just as importantly, it is easier to find the
ROIs. Quantification of mBC in the ROI [see Fig. 6(c)] is able to
reflect the presumed general time course showing the expected
increase and decrease of the QPC during the preictal, ictal, and
postictal periods. The increase of time smoothing in the MWT
approach [see (A) to (B)] is clearly visible, time smoothing of
MGT through the use of different b ounds [see (C) to (D)] is
less distinct.
Statistical analysis of mBC courses over time confirms regions with significantly increased mBC [e.g., before 100, at
240, at 300, after 400 s; red line in Fig. 6(c)].
Confidence tubes of mBC achieved by the bootstrapping approach supports this view and provides more insight into statistical properties of mBC over time. Confidence intervals at,
e.g., 240 s (preictal period), were much narrower than after the
seizure (postictal period), where the narrowest confidence intervals appear during the seizure [ictal period; Fig. 6(c)]. Best
combination of significantly increased mBC (red line) and discrimination between different mBC periods over time (confidcence tubes) can be achieved by the second MGT approach [see
Fig. 6(c) and (D)].
2) EEG of Neonates With BurstInterburst Pattern During
QS: Mean results of one neonate (#1; K = 17 EEG trials) of
MWT and MGT-based tvPS of burstinterburst pattern in the
EEG are depicted in Fig. 7(a), time courses of mBC in the
ROI are given in Fig. 7(b). Grand mean results of all neonates
(K = 102 EEG trials) using MWT and MGT are not shown but
confirm the results of the single neonate.
The frequency-dependence of TFR in the tvPS of MWT is
clearly visible [see Fig. 7(a-a) and (a-b)]. Again, the frequency
resolution in MWT-based approach (A) is not high enough,
whereas approach (B) is afflicted with strong time smoothing
(low time resolution). The TFR of all MGT-based approaches
[see Fig. 7(a-c) and (a-d)] is adequate for bispectral analysis.
Quantification of mBC in the ROI [see Fig. 7(b)] is able to reflect
the presumed general time course of increased QPC with the
onset of the burst period. Different TFR (and resulting different
time smoothing) of the MWT approaches [see (A) to (B)] is
clearly visible, as well as the time smoothing of MGT through
the use of different b ounds [see (C) to (D)]. Statistical analysis
of the mBC course over time confirms regions with significantly
increased mBC after the burst onset [e.g., at 5, 6, 8 s; red line in
Fig. 7(b)].
The course of the mBC confidence tubes achieved by the
bootstrapping approach shows statistical properties of mBC over
time. Best combination of significantly increased mBC (red

SCHIECKE et al.: MATCHING PURSUIT-BASED TIME-VARIANT BISPECTRAL ANALYSIS AND ITS APPLICATION TO BIOMEDICAL SIGNALS

1945

Fig. 6. Grand mean results (K = 18) of HRV during TLE. (a) tvPS in (b1), (b2), and (b3) BA at distinct time points (240, 340, and 390 s; see Fig. 3(a), and
in (c) mBC in the ROI are shown. (A) and (B) depict results of MWT-based approaches ( 0 = 2/10). (C) and (D) depict results of MGT-based approaches
( b o u n d s = [5 s, Inf]/[10 s, Inf]). Red color designates high power, blue color: low power of tvPS or BA, respectively, in columns (a), (b1), (b2), and (b3). In
column (c), time course of mBC in the ROI = [0.075Hz, 0.15Hz] [0.25Hz, 0.35Hz] is shown. Choice of ROI for the mBC quantification is illustrated by a
gray frame in (b1) to (b3). Red lines in column (c) designate 5% threshold for statistical significance of mBC, gray filled areas indicate 95% confidence tubes of
mBC achieved by a bootstrapping approach.

line) and discrimination between different mBC periods over

time (confidcence tubes, before and after burst onset) can be
achieved by the second MGT approach [see Fig. 7(b) and (D)].

V. DISCUSSION AND SUMMARY

Fig. 7. Mean results (K = 17) of burstinterburst pattern in the EEG of

one child (#1). (a) tvPS and in (b) time course of mBC in the ROI =
[0.5 Hz, 1.5 Hz] [3 Hz, 5 Hz] is shown. (A) and (B) depict the results of
MWT-based approaches ( 0 = 2/5). (C) and (D) depict results of MGTbased approaches ( b o u n d s = [0.1 s, Inf]/[0.25 s, Inf]). Gray line in column
(b) indicates the onset of burst, red lines designate 5% threshold for statistical
significance of mBC, gray filled areas indicate 95% confidence tubes of mBC
achieved by a bootstrapping approach.

The methodological novelty of our study is that for the first

time, MP decomposition has been used to achieve a time-variant
bispectral analysis without having to perform a manual (experimental) parameter adjustment for optimizing TFR, as required
in GT-, MWT-, and AR-based bispectral analysis. To emphasize
the advantage of our approach, the TFR properties of our method
and the more common approaches has to be compared. The
STFT and the GT are characterized by a frequency-independent
TFR, i.e., all frequency components of the signal are analyzed
with identical TFRs. MWT-based methods have a frequencydependent TFR, where LF components are analyzed with HF
and low time resolution and HF components with LF and high
time resolution. This uncertainty principle is valid for all timevariant methods. This was illustrated in an excellent manner by
Jamsek et al. [6]. The authors exposed the specific problems
by a comparison of STFT- and MWT-based bispectral analysis.
The time and the frequency resolution of third-order TFDs and
the bispectral approaches based on them can be adjusted independently. This does not conflict with the uncertainty principle
because the resolutions are limited by the so-called local moment function. Additionally, non-QPC terms in the bispectrum
occur which can be sufficiently suppressed by a broadening of
the time window, which causes a decrease in the time resolution,
i.e., a compromise between both properties must be found. The

1946

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

TFR of time-variant parametric approaches depends on the order

of the AR model (frequency resolution) and (smoothing) parameters of the estimation algorithm (time resolution). The lower
the model order (number of AR parameters =
window length of
the AR process analyzer), the higher the time resolution is and
the more smoothed the spectrum is (LF resolution). If the model
order is too high, then spurious peaks occur [26]. The model
order must be estimated by estimation methods which stationary approaches commonly use. Smoothing during time-variant
AR parameter estimation improves the estimation properties but
decreases the time resolution. In practice, an acceptable compromise needs to be found in order to obtain the most appropriate
TFR. Such an optimization procedure was described in detail
by Schwab et al. [27].
To overcome such limitations characteristic to TFR, the recently introduced MGT is used [14], which consists of a combination of the MP and the GT providing a complex time
frequency plane with an optimal TFR for each signal component. The superiority of the MGT has already been demonstrated
in linear-phase locking and synchronization analyses for comparison of the GT- and MWT-approaches [14]. The results of
this study also demonstrate that the application of the MGT has
beneficial effects for QPC analysis.
Kus et al. [28] concluded that the limited acceptance of MP
in the EEG analysis may be partly due to the lack of defined
criteria (rules) for setting the most important parameters of the
algorithm, which are the number and distribution of atoms. In
general, the results of MP-based processing concepts depend to a
large extent on the dictionary used. The MGT approach is based
on a Gabor atom dictionary, which allows the representation
of the signal into the complex timefrequency plane, i.e., both
time-variant amplitude and phase information are preserved.
The precondition is that the natural signal components can be approximated by Gabor atoms, which is often, but not always, the
case in HRV and EEG. Three further aspects with regard to the
dictionary should be discussed. First, both signals can contain
asymmetric components, which cannot be reliably described by
Gabor atoms. To detect and quantify such components, a dictionary which is enriched by or composed of asymmetric atoms
can be used [29]. Transient HRV patterns (e.g., extrasystoles and
other cardiac arrhythmias) and artifacts (incorrect QRS triggering) are mostly asymmetric, and they can mask or disturb the
HRV analysis. Therefore, MP with asymmetric dictionaries can
be used for the detection of disturbing HRV patterns. Second,
if the atoms from the dictionary are shifted on a discrete time
grid to cover the whole time range, the discrete positions automatically introduce bias. To compensate for this, the so-called
stochastic dictionaries can be used, where the decomposition is
carried out for many different, randomly perturbed dictionaries [30]. Third, the size of the dictionary plays an important
role. In general, the use of a larger dictionary of atoms for
decomposition of the same signal should lead to more accurate
parametrization [28]. Kus et al. [28] introduced an approach
for the design of an optimal Gabor dictionary, which is available
by an open-source software package. These advancements can
be considered for further methodological improvement of our
approach.

For MGT-based QPC analysis, the use of a lower sigma bound

for the GT is advantageous. Our AM simulation study (AAM
and MAM) shows that AM-related QPC can be successfully detected by using this thresholding operation for time resolution,
i.e., Gabor atoms of the dictionary with sigma values below
this bound (with time resolution, which is too high) will be
computed (GT) with a fixed (bound value) lower time resolution. This is more appropriate for the particular characteristics
of the AM signal, i.e., the carrier and the side-band frequencies occur as continuous traces in the timefrequency plane.
AM phenomena have been already described in HRV and EEG
(e.g., [20] and [21]) and have been investigated in envelopeto-signal-correlation approaches [31], which are based on AM
demodulation (envelope) [32]. In contrast, bispectral analysis
is a more general approach. Bispectral analysis is sensitive to
changes in third-order statistics, i.e., any nonlinearity that alters the (third order) statistics of the input will be detectable by
it [33].
In biomedical signal analysis, such statistical changes are
certainly an indication of a change of state within a physiological
system or process. Our HRV and EEG data show abrupt changes
in state, and we have shown that the MGT-based bispectral
analysis highlights the QPC changes associated with this more
clearly than the MWT approaches.
Our methodological approach is highly relevant to physiological and clinical questions. RSA or HF HRV, which can be defined as an HRV component of the frequency of breathing [34], is
often used as an index of cardiac vagal tone, whereby the HR
increases during inspiration and decreases during expiration
[35]. RSA amplitude is modulated by different physiological
parameters (depth and frequency of breathing, cardiac vagal efferent effects etc.; for an overview, see [36]). The rhythmic AM
of the RSA is one possible cause for QPC in HRV, e.g., a modulation by the LF waves of the HRV (e.g., by the 0.1-Hz waves).
It can be speculated that the nonlinearity of the cardiovascular
cardiorespiratory control system contributes to the occurrence
of QPC and that a change in the physiological state is accompanied by a change of the control behavior. In clinical practice,
timefrequency features of HRV are frequently used, e.g., for
an automatic detection of seizure in newborns [37]. The results
of our study confirm our previous results that the QPC increases
before and, in particular, after an epileptic seizure (temporal lobe
epilepsy). A neurophysiological link between the 0.1-Hz HRV
waves and the Mayer waves in systemic arterial blood pressure
exists. Mayer waves are correlated with the oscillations of efferent sympathetic nervous activity and the baroreflex plays a
major role in their generation. In contrast, the 0.1-Hz HRV component associated with the Mayer wave includes most probably
both sympathetic and parasympathetic (vagal) influences [19].
Our results indicate a changed coordination between LF and
HF HRV waves in particular after the seizure, i.e., the autonomic nervous system still remains in an altered physiological
state.
As has been stated, The nonlinear nature of neuronal activity
contributes to the formation of an EEG signal with very complex
dynamics [38] and as has been shown, QPC between different
frequency components of EEG aids in discovering changes in

SCHIECKE et al.: MATCHING PURSUIT-BASED TIME-VARIANT BISPECTRAL ANALYSIS AND ITS APPLICATION TO BIOMEDICAL SIGNALS

physiological states of the brain and helps to elucidate the complex dynamics of physiological systems. For example, bispectral analysis has been used for the investigation of EEG patterns
following hypoxic-asphyxic arrest [39] for the identification of
epileptic and focal ischemic cerebral EEG [40], [41] as well as
to detect sleep [42], anesthesia [43], and sedation states [5]. Our
proof-of-principle application in the field of EEG analysis is
related to QPC changes in neonatal EEG patterns derived from
the quiet sleep state. We used the EEG of mature neonates to
demonstrate the methodological improvements by using MGT.
After the burst onset, the bicoherence increases, which indicates
a trigger process in which most probably the thalamus is initially
involved [44]. After the burst onset, the cortical LF oscillation
modulates the amplitude of HF oscillatory activities. This can
be physiologically explained by the model of Steriade [45], in
which a depolarization phase of a cortical LF oscillation travels through the corticothalamic pathway and triggers a spindle
sequence in the reticular thalamic nucleus that will be delivered to the cortex via the dorsal thalamus. We have previously
shown that changes in the QPCs degree after the burst onset are
associated with brain maturation [23].
For bispectral analysis, i.e., independent from the computation method, defining the ROI for the appropriate selection of
parameters from the bispectral plane is important and is based
on a priori information. The optimal TFR of the MGT approach
is helpful for this decision making because the real signal structure is mapped in the timefrequency plane. This is also true for
coupling structures in the bispectral plane. Therefore, bispectral ROIs can be better identified, e.g., ROIs which enable the
investigation of specific physiological signal components and
their couplings. This has been shown by our simulation study
(individually tailored versus guideline-driven ROIs; red lines in
Fig. 4(D)).
VI. CONCLUSION
In general, it can be concluded that the application of the new
MGT approach leads to the improved QPC results because of the
achieved optimal TFR, the use of sigma bounds, and thereby, the
resulting advanced decision making in terms of defining specific
estimation parameters. The next question arises, how can these
advantages be optimally expanded for use in further research
of for application in the clinical setting? One concept is to use
dictionary modifications or dictionary alternatives to improve
methodologies. It can be expected that the advancements in
MP and application-driven MP adaptations will lead to further
improvement of time-variant bispectral analysis in biomedicine.
REFERENCES
[1] C. L. Nikias and A. P. Petropulu, Higher-Order Spectra Analysis. A Nonlinear Signal Processing Framework. PTR Prentice Englewood Cliffs, NJ,
USA: Prentice-Hall, 1993.
[2] K. C. Chua et al., Application of higher order statistics/spectra in biomedical signals-A review, Med. Eng. Phys., vol. 32, pp. 679689, Sep. 2010.
[3] P. J. Huber et al., Statistical methods for investigating phase relations
in stationary stochastic processes, IEEE Trans. Audio Electroacoustics,
vol. AU-19, no. 1, pp. 7886, Mar. 1971.

1947

[4] V. Chandran, Time-varying bispectral analysis of visually evoked multichannel EEG, Eurasip J. Adv. Signal Process., vol. 2012, Jul. 12, 2012.
[5] H. Witte et al., Analysis and modeling of time-variant amplitudefrequency couplings of and between oscillations of EEG bursts, Biol.
Cybern., vol. 99, pp. 139157, Aug. 2008.
[6] J. Jamsek et al., Wavelet bispectral analysis for the study of interactions
among oscillators whose basic frequencies are significantly time variable,
Phys. Rev. E, vol. 76, Oct. 2007.
[7] J. R. Fonollosa and C. L. Nikias, Wigner higher-order moment spectra
- definition, properties, computation and application to transient signal
analysis, IEEE Trans. Signal Process., vol. 41, no. 1, pp. 245266, Jan.
1993.
[8] M. Helbig et al., Analysis of time-variant quadratic phase couplings in the
trace alternant EEG by recursive estimation of 3rd-order time-frequency
distributions, J. Neurosci. Meth., vol. 157, pp. 168177, Oct. 15,
2006.
[9] B. Boashash et al., Time-frequency processing of nonstationary signals,
IEEE Signal Process. Mag., vol. 30, no. 6, pp. 108119, Nov. 2013.
[10] K. Schwab et al., Time-variant parametric estimation of transient
quadratic phase couplings during electroencephalographic burst activity,
Meth. Inform. Med., vol. 44, pp. 374383, 2005.
[11] G. Valenza et al., Point-process nonlinear autonomic assessment of
depressive states in bipolar patients, Meth. Inform. Med., vol. 53,
pp. 296302, 2014.
[12] G. Valenza et al., Point-process nonlinear models with Laguerre and
Volterra expansions: Instantaneous assessment of heartbeat dynamics,
IEEE Trans. Signal Process., vol. 61, no. 11, pp. 29142926, Jun. 2013.
[13] M. Wacker and H. Witte, Time-frequency techniques in biomedical signal
analysis. a tutorial review of similarities and differences, Meth. Inform.
Med., vol. 52, pp. 279296, 2013.
[14] M. Wacker and H. Witte, Adaptive phase extraction: Incorporating the
gabor transform in the matching pursuit algorithm, IEEE Trans. Biomed.
Eng., vol. 58, no. 10, pp. 28442851, 2011.
[15] S. G. Mallat and Z. F. Zhang, Matching pursuits with timefrequency dictionaries, IEEE Trans. Signal Process., vol. 41, no. 12,
pp. 33973415, Dec. 1993.
[16] J. Theiler et al. Testing for nonlinearity in time-seriesthe method of
surrogate data, Physica D, vol. 58, pp. 7794, Sep. 15, 1992.
[17] D. Piper et al., Time-variant coherence between heart rate variability and
EEG activity in epileptic patients: an advanced coupling analysis between
physiological networks, New J. Phys., vol. 16 p. 115012, 2014.
[18] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap. Boca
Raton, FL, USA: CRC Press, 1994.
[19] A. J. Camm et al., Heart rate variability. Standards of measurement,
physiological interpretation, and clinical use, Eur. Heart J., vol. 17,
pp. 354381, Mar. 1996.
[20] H. Witte et al., Technique for the quantification of transient quadratic
phase couplings between heart rate components, Biomedizinische Technik, vol. 46, pp. 4249, Mar. 2001.
[21] D. B. Chorlian et al., Amplitude modulation of gamma band oscillations
at alpha frequency produced by photic driving, Int. J. Psychophysiol.,
vol. 61, pp. 262278, Aug. 2006.
[22] K. Schiecke et al., Time-variant, frequency-selective, linear and nonlinear
analysis of heart rate variability in children with temporal lobe epilepsy,
IEEE Trans. Biomed. Eng., vol. 61, no. 6, pp. 17981808, Jun. 2014.
[23] M. Wacker et al., A processing scheme for time-variant phase analysis
in EEG burst activity of premature and full-term newborns in quiet sleep:
a methodological study, Biomed. Eng.-Biomedizinische Technik, vol. 57,
pp. 491505, Dec. 2012.
[24] I. Daskalov and I. Christov, Improvement of resolution in measurement
of electrocardiogram RR intervals by interpolation, Med. Eng. Phys.,
vol. 19, pp. 375379, Jun. 1997.
[25] A. S. French and A. V. Holden, Alias-free sampling of neuronal spike
trains, Kybernetik, vol. 8, pp. 165171, 1971.
[26] J. G. Proakis and D. G. Manolakis, Introduction to Digital Signal Processing. New York, NY, USA: Macmillan, 1989.
[27] K. Schwab et al., Time-variant parametric estimation of transient
quadratic phase couplings between heart rate components in healthy
neonates, Med. Biol. Eng. Comput., vol. 44, pp. 10771083, Dec. 2006.
[28] R. Kus et al., Multivariate matching pursuit in optimal Gabor dictionaries:
theory and software with interface for EEG/MEG via Svarog, Biomed.
Eng. Online, vol. 12, Sep. 23, 2013.
[29] C. Baumgartner et al., Discussion of time-frequency techniques in
biomedical signal analysis: A tutorial review of similarities and differences, Meth. Inform. Med., vol. 52, pp. 297307, 2013.

1948

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 62, NO. 8, AUGUST 2015

[30] P. J. Durka et al., Stochastic time-frequency dictionaries for matching

pursuit, IEEE Trans. Signal Process., vol. 49, no. 3, pp. 507510, Mar.
2001.
[31] A. Bruns and R. Eckhorn, Task-related coupling from high- to lowfrequency signals among visual cortical areas in human subdural recordings, Int. J. Psychophysiol., vol. 51, pp. 97116, Jan. 2004.
[32] M. Arnold et al., Time-variant investigation of quadratic phase couplings
caused by amplitude modulation in electroencephalic burst-suppression
patterns, J. Clin. Monit. Comput., vol. 17, pp. 11151123, Feb. 2002.
[33] J. M. Nichols et al., The bispectrum and bicoherence for quadratically
nonlinear systems subject to non-Gaussian inputs, IEEE Trans. Signal
Process., vol. 57, no. 10, pp. 38793890, Oct. 2009.
[34] A. Ben-Tal et al., Central regulation of heart rate and the appearance of
respiratory sinus arrhythmia: New insights from mathematical modeling,
Math. Biosci., vol. 255, pp. 7182, Sep. 2014.
[35] A. Ben-Tal et al., Evaluating the physiological significance of respiratory
sinus arrhythmia: Looking beyond ventilation-perfusion efficiency, J.
Physiol., London, vol. 590, pp. 19892008, Apr. 2012.
[36] P. Grossman and E. W. Taylor, Toward understanding respiratory sinus
arrhythmia: Relations to cardiac vagal tone, evolution and biobehavioral
functions, Biol. Psychol., vol. 74, pp. 263285, Feb. 2007.
[37] M. B. Malarvili and M. Mesbah, Newborn seizure detection based
on heart rate variability, IEEE Trans. Biomed. Eng., vol. 56, no. 11,
pp. 25942603, Nov. 2009.
[38] J. C. Sigl and N. G. Chamoun, An introduction to bispectral analysis for
the electroencephalogram, J. Clin. Monit., vol. 10, pp. 392404, Nov.
1994.

[39] J. Muthuswamy et al., Higher-order spectral analysis of burst patterns

in EEG, IEEE Trans. Biomed. Eng., vol. 46, no. 1, pp. 9299, Jan.
1999.
[40] K. C. Chua et al., Application of higher order spectra to identify epileptic
EEG, J. Med. Syst., vol. 35, pp. 15631571, Dec. 2011.
[41] J. W. Zhang et al., Bispectrum analysis of focal ischemic cerebral EEG
signal using third-order recursion method, IEEE Trans. Biomed. Eng.,
vol. 47, no. 3, pp. 352359, Mar. 2000.
[42] X. L. Li et al., The comodulation measure of neuronal oscillations with
general harmonic wavelet bicoherence and application to sleep analysis,
Neuroimage, vol. 48, pp. 501514, Nov. 15, 2009.
[43] K. Hayashi et al., Simultaneous bicoherence analysis of occipital and
frontal electroencephalograms in awake and anesthetized subjects, Clin.
Neurophysiol., vol. 125, pp. 194201, Jan. 2014.
[44] H. Witte et al., Time-variant analysis of phase couplings and amplitudefrequency dependencies of and between frequency components of EEG
burst patterns in full-term newborns, Clin. Neurophysiol., vol. 122,
pp. 253266, Feb. 2011.
[45] M. Steriade, Grouping of brain rhythms in corticothalamic systems,
Neuroscience, vol. 137, pp. 10871106, 2006.

Authors photographs and biographies not available at the time of publication.