Bi Spectrum For EEG Analysis

AN INTRODUCTION TO THE BISPECTRUM FOR EEG ANALYSIS
D. Wong, D.A. Clifton, L. Tarassenko
Biomedical Signal Processing Group, Institute of Biomedical Engineering, University of Oxford, Oxford, UK
wong@robots.ox.ac.uk
Abstract - This paper provides a tutorial for bispectral

analysis, a signal processing technique commonly used is the triple product. Using the convolution
for the analysis of the Electroencephalogram (EEG). theorem for Fourier transforms, the bispectrum
The use of this technique has been hindered by popular
misconceptions deriving from existing tutorial papers. of a process, x, is formally defined as:
INTRODUCTION (3)
Analysis of EEG is typically performed where c3(τ1,τ2)=E{xt,xt+τ1,xt+τ2}, and E{.} is the

using Fourier analysis, which is useful for expectation operator.
detecting frequency components that Although other methods of calculating the
correspond to the mental state of a patient. bispectrum are also used in practice, the
Typically, such information is obtained from examples below use equation 1 for the sake of
the magnitude spectrum, as the information clarity.
content of the phase spectrum can be harder to
interpret. A. Quadratically Phase Coupled Components
Although use of the magnitude spectrum is Three sinusoid components, x1,x2 and x3, of
sometimes sufficient, in many instances, the a signal are quadratically phase coupled (QPC)
phase information is critically important. if the frequency and phase of the x3
Bispectral analysis offers a way of gaining component is equal to the sum of the
further phase information by detecting phase frequencies and phases of the x1 and x2
relationships between frequency components. components. For example, for
Such relationships have been shown to exist in
a medical context during periods where a the components are
patient has an impaired mental state (for QPC only if f1+f2=f3 and
example see [1]).
The rest of this paper outlines the
mathematical principles behind the bispectrum,
and provides simple examples that demonstrate
its use. Using the knowledge gained from the
tutorial, we will dispel a commonly-held
misconception.
II. BISPECTRUM TUTORIAL
Bispectral analysis makes use of phase

information by detecting whether the phase of
signal components at frequencies , , and
are interdependent. Furthermore, the degree of
dependence is quantified so that a high
bispectrum correlates to a signal with highly
interdependent frequency components.
Knowing this, it makes intuitive sense that
the bispectrum can be calculated from using
the Fourier transform of a signal evaluated at
, , and . The bispectrum of x can be
expressed mathematically as Figure 1: A graphical interpretation of the
(1) bispectrum calculation for a synthetic QPC signal.
The signal is shown above its three component
where X denotes the Fourier transform of x. In sinusoids.
this case, the signal has been split into k time
segments that we call epochs, and A simple QPC signal, along with its
component sinusoids, is shown in Fig. 1. The
(2) signal is split into k=3 epochs (where the third
epoch is truncated), and the component III. DISCUSSION
frequencies and phases are calculated at the
start of each epoch. The triple products can This tutorial has demonstrated that the
then be calculated using eq. 2: bispectrum can be used to identify phase
coherence between any three sinusoids that are
components of a signal. The coherence is
detected for QPC components (example A), for
coherent, but non-QPC components (example
B), but not for incoherent signals where the
relative phase between two fundamentals
The mean of the triple products gives the changes over time (example C).
estimate of the bispectrum. In this case, Previous tutorial papers on the bispectrum,
B(f1,f2)=π3 including those by Sigl and Chamoun [2],
Nikias and Raghuveer [3], Shen [4], and Holt
B. Non-Quadratic Coupling [5], have asserted that the technique is only
Figure 2 shows a similar system, but where valid for QPC signals. They each give a simple
the components are no longer QPC, as system such as:
. However, the phases are
coherent, meaning that the relative phase
between any of the components does not )
change over time. Triple products are
calculated as before, resulting in |B(f1,f2)|=π3. and apply theconstraints ,
, , , and
are random and independent so
that the mixture is not QPC.
They go on to state that such a system will
result in a zero bispectrum over all the
frequency space. However, this assertion is
incorrect and inconsistent with example B.
In both example B and the system described
above, the phases are set to be random
constants. Thus, the signal components are
coherent as there is no relative phase drift
between them, and the bispectrum will be
significant.
In conclusion, we have sought to provide an
intuitive guide to the bispectrum using a
graphical explanation, and have proceeded to
show that previous literature on the subject is
confusing and misleading.
REFERENCES
Figure 2: A graphical interpretation of the
bispectrum calculation for a non-QPC, coherent, [1] J Jeong. EEG dynamics in patients with
synthetic waveform. Alzheimer's disease. Clin. Neurophysiology,
115:1490-1505, 2004
C. Incoherent components [2] JC Sigl and NG Chamoun. An introduction to
The third example, which is not pictured here bispectral analysis for the electroencephalogram. J.
for brevity, assumes a non-QPC and non- Clin. Monitoring and Comp.10(6):392-404, 1994
coherent set of component sinusoids. By non- [3] CL Nikias and MR Raghuveer. Bispectrum
estimation: a digital signal processing framework.
coherent, it is meant that the relative phases
Proc. of IEEE, 75(7):869-891, 1987
between the sinusoids change over time, so [4] M Shen, F Chan, L Sun, P Beadle. Parametric
are time-dependent variables. bispectral estimation of EEG signals in different
In this case, the estimate of the bispectrum is functional states of the brain. IEE Proc. Science,
once again calculated from equation 2 to give measurement and tech. 147(6):374-377,2000
[5] M Holt. The use of neural networks in the
. The magnitude of the bispectrum
analysis of the anaesthetic EEG. PhD thesis,
is much less than in the other examples. In U.Oxford, 1997
fact, for incoherent signals, the bispectrum
tends toward zero when a larger number of
epochs are used.

Bi Spectrum For EEG Analysis

Uploaded by

Copyright:

Available Formats

Bi Spectrum For EEG Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bi Spectrum For EEG Analysis

Uploaded by

Copyright:

Available Formats

AN INTRODUCTION TO THE BISPECTRUM FOR EEG ANALYSIS

D. Wong, D.A. Clifton, L. Tarassenko

Abstract - This paper provides a tutorial for bispectral

Analysis of EEG is typically performed where c3(τ1,τ2)=E{xt,xt+τ1,xt+τ2}, and E{.} is the

II. BISPECTRUM TUTORIAL

Bispectral analysis makes use of phase

You might also like