=Paper=
{{Paper
|id=Vol-1638/Paper82
|storemode=property
|title=Development of a method of analysis of EEG wave packets in early stages of Parkinson's disease
|pdfUrl=https://ceur-ws.org/Vol-1638/Paper82.pdf
|volume=Vol-1638
|authors=Olga S. Sushkova,Alexey A. Morozov,Alexandra V. Gabova
}}
==Development of a method of analysis of EEG wave packets in early stages of Parkinson's disease ==
https://ceur-ws.org/Vol-1638/Paper82.pdf
Mathematical Modeling
DEVELOPMENT OF A METHOD OF ANALYSIS OF
EEG WAVE PACKETS IN EARLY STAGES OF
PARKINSON'S DISEASE
O.S. Sushkova1, A.A. Morozov1,2, A.V. Gabova3
1
Kotel'nikov Institute of Radio Engineering and Electronics of RAS, Moscow, Russia
2
Moscow State University of Psychology & Education, Moscow, Russia
3
Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow, Russia
Abstract. A method of analysis of EEG wave packets based on wavelets and
nonparametric statistics is developed. The method is compared with standard
methods based on Fourier spectra and complex Morlet wavelets by the example
of Parkinson's disease experimental data. We demonstrate that these methods
are complementary, that is, the standard methods and the wave packet analysis
method reveal sufficiently different effects in the EEG data.
Keywords: wave packet, wave train, burst, electroencephalogram, EEG, wave-
let, nonparametric statistics, decrease of quantity of wave packets, increase of
alpha power spectral density.
Citation: Sushkova OS, Morozov AA, Gabova AV. Development of a method
of analysis of EEG wave packets in early stages of Parkinson's disease. CEUR
Workshop Proceedings, 2016; 1638: 681-690. DOI: 10.18287/1613-0073-2016-
1638-681-690
Introduction
A wave packet is a wave action (a “burst”) that is well localized in space and time.
The wave packet is a typical pattern in a background electroencephalogram (EEG)
and detecting / analysing such signals gives useful information about the brain activi-
ty. Alpha spindles (sleep spindles) and beta spindles are the best known examples of
the wave packets in EEG; several methods based on Fourier spectra, wavelets, auto-
regressive models, adaptive filtering, etc. have been developed for detecting and ana-
lysing these EEG patterns (see surveys in [1,2,3]).
The idea of our method of EEG analysis is in that we detect and analyse the wave
packets in a wide frequency band including theta, alpha, beta, and gamma EEG. We
detect the wave packets as local maxima in a wavelet spectrogram of an EEG record
and compute various attributes of these signals: a quantity of the wave packets per
second, an average power / amplitude, an average central frequency, an average dura-
tion of the signals, etc. Then we implement a group statistical analysis on the base of
Information Technology and Nanotechnology (ITNT-2016) 681
Mathematical Modeling Sushkova OS, Morozov AA. et al…
these attributes using the nonparametric statistics. In the paper, we demonstrate that
this method reveals a new effect in the early stages of Parkinson's disease.
The method of analysis of the EEG wave packets is described in Section “The Analy-
sis Method”. The experimental setting is described in Section “The Experimental
Setting”. The data processing stages and results of the statistical analysis are de-
scribed in Section “Data Analysis”. Section “Method Comparison” contains a com-
parison of the method with the standard methods based on Fourier spectra and com-
plex Morlet wavelets. Section “Discussion” contains a discussion of the results of the
statistical analysis and the comparison.
The Analysis Method
The method of EEG wave packet analysis includes detection of local maxima in
wavelet spectrograms, determination of various attributes of these maxima, and statis-
tical analysis of these attributes.
Let M be a local maximum in a wavelet spectrogram. We consider M as a case of a
wave packet if the half width of M (at half maximum) is greater or equal to the TH
threshold in the time plane and greater or equal to the FH threshold in the frequency
plane (see fig. 1). The FH threshold is a constant and the TH threshold is a function of
the f central frequency of the M maximum:
TH = NP / (2f),
where NP is a constant given by an expert. In this paper, we apply the values: NP=2
and FH=1 Hz.
Fig. 1. An example of a spectrogram of a wave packet in a time-frequency domain. The dia-
gram on the left shows the spectrogram of the signal in the time plane; the abscissa indicates a
time and the ordinate indicates a power. The diagram on the right shows the spectrogram of the
signal in the frequency plane; the abscissa indicates a frequency and the ordinate indicates a
power
Let us consider a case of the wave packet analysis by the example of Parkinson's dis-
ease experimental data (fig. 2).
The amplitude of a typical wave packet in the alpha frequency band 8-12 Hz (an alpha
spindle) is much bigger than the amplitude of a typical wave packet in the beta fre-
quency band 12-25 Hz (a beta wave packet). Thus the beta wave packets are hardly
noticeable against a background of the alpha wave packets. Moreover, spectrograms
Information Technology and Nanotechnology (ITNT-2016) 682
Mathematical Modeling Sushkova OS, Morozov AA. et al…
of alpha spindles have tails that continue in other frequency bands including the beta
and the gamma bands. Thus classical EEG analysis methods based on wavelets can
mistakenly account these tails of the alpha wave packets as a beta / gamma electrical
activity.
Fig. 2. A wavelet spectrogram of a background EEG record of an early stage Parkinson's dis-
ease patient. One can observe wave packets in the alpha frequency range 8-12 Hz well, but
wave packets in the beta frequency range are hardly noticeable, because they are much smaller
than the alpha wave packets
The method under consideration detects a set of wave packets in various frequency
bands (fig. 3) in the complex Morlet spectrogram (fig. 4). Let us consider a spectro-
gram of a wave packet B in a time-frequency domain (fig. 3) and a neighbourhood of
the B wave packet in the wavelet spectrogram (see fig. 5).
Fig. 3. A set of EEG wave packets in the time-frequency domain. The B wave packet is indicat-
ed by an arrow
The B wave packet has a distinct localization in the beta time-frequency area. At the
same time, one can see a pronounced tail of another local maximum A on the left that
is originated in the alpha frequency band. The developed method prevents erroneous
recognition of this tail as an electrical activity in the beta band. The original wave
packet B is shown in figure 6.
Information Technology and Nanotechnology (ITNT-2016) 683
Mathematical Modeling Sushkova OS, Morozov AA. et al…
Fig. 4. An example of a wavelet spectrogram of an EEG record of an early stage Parkinson's
disease patient (a view from above). The abscissa indicates a time (the 0-150 sec interval) and
the ordinate indicates a frequency (in Hz)
Fig. 5. A spectrogram of the B wave packet in the beta band. One can see a pronounced tail of
another local maximum A on the left that is originated in the alpha band
Fig. 6. The original EEG signal. The B wave packet is indicated by an ellipse with an arrow
At the next stage of the analysis, a set of attributes of detected wave packets is calcu-
lated including the following ones:
The quantity of the wave packets per second in given frequency bands.
An average power / amplitude of the wave packets in the given bands.
An average frequency of the wave packets in the given frequency bands.
Information Technology and Nanotechnology (ITNT-2016) 684
Mathematical Modeling Sushkova OS, Morozov AA. et al…
A standard deviation of the power / amplitude of wave packets in the given fre-
quency bands.
An average duration of the wave packets in the given frequency bands.
Then the Mann-Whitney nonparametric statistical test is used for group comparisons
of the calculated values. A nonparametric statistics is necessary because we analyze
samples of small sizes (one can compare 15 subject groups and even less) and the
samples do not correspond to the normal distribution law.
The Experimental Setting
A group of 18 patients with the first stage Parkinson’s disease receiving no treatment
and a group of 19 healthy volunteers as a control were recruited. The patients were
diagnosed according to the standard Hoehn and Yahr scale. All patients and volun-
teers were right-handers. There were 11 patients with the right side tremor and 7 pa-
tients with the left side tremor in the group. The age of patients ranged from 38 to 64
years; the mean age was 55 years; the standard deviation of the age was 6.5 years; the
median of the age was 56 years; the 10% quantile of the age was 47.30; the 90%
quantile was 62 years. The ages of controls ranged from 40 to 81 years; the mean age
was 55 years; the standard deviation of the age was 10.51 years; the median was 53
years; the 10% quantile was 43.40 years; the 90% quantile was 72.60 years. No statis-
tically significant differences between the ages of the patients and the controls were
detected.
EEG was recorded in a non-standard condition [4], that is, a subject was instructed to
keep a special pose to provoke a tremor: the arms were placed on armrests of the
chair; the palms were straightened, placed in a vertical plane, and stretched a bit; the
feet were stretched a bit and touched the floor by the heels only. The eyes were closed
during the recording.
A 41-channel digital EEG system Neuron-Spectrum-5 (Neurosoft Ltd.) was used for
the data acquisition. The sampling rate was 500 Hz. The 0.5 Hz high-pass filter and
the 50 Hz supply-line filter were used. Three EEG records were acquired for every
subject with interruptions for a rest and a relaxation. The duration of every record was
not less than 2 minutes. Then the best of three records was selected that contains a
minimal number of artefacts.
A standard 10x20 EEG acquisition schema was used. In this paper, the C3 and C4
electrodes are considered only, because these electrodes approximately correspond to
the motor cortex areas and are situated in the scalp area that produces a minimal num-
ber of muscle artefacts.
Data Analysis
Special software was developed for analyzing the data. The analysis includes the fol-
lowing EEG pre-processing operations:
Information Technology and Nanotechnology (ITNT-2016) 685
Mathematical Modeling Sushkova OS, Morozov AA. et al…
The Huber's X84 method [5] for outlier rejection was used for removing EEG arte-
facts.
A set of notch filters was applied for removing a power-line noise at 50, 100, 150,
and 200 Hz.
The eight order 2-240 Hz band pass Butterworth filter was applied. Signals were
filtered in the forward and reverse directions to eliminate a phase distortion.
Signals were decimated with the decimation factor 8.
The spectrograms were created using the Morlet wavelet:
x2
exp 2iFc x exp
1
( x)
Fb Fb
In this paper, Fb equals 1 and Fc equals 1. The frequency step in the spectrograms
equals 0.1 Hz.
The method of EEG wave packet analysis reveals a new effect in the Parkinson's dis-
ease, namely, the number of beta (12-25 Hz) wave packets in the C3 and C4 channels
is significantly decreased (Mann-Whitney, p < 0.02), see figure 7. Note that the quan-
tity of the wave packets only is considered in this test, but not the amplitude of the
wave packets.
Fig. 7. Scattering of the quantity of beta wave packets in patients and controls. The C3 elec-
trode is shown at the left, the C4 electrode is shown at the right. The abscissa is the number of
wave packets per second in the alpha frequency band. The ordinate is the number of wave
packets in the beta frequency band. The patients are indicated by diamonds and the controls are
indicated by circles
Information Technology and Nanotechnology (ITNT-2016) 686
Mathematical Modeling Sushkova OS, Morozov AA. et al…
Method Comparison
Let us compare the developed method of EEG wave packet analysis with a standard
method of EEG analysis based on the Welch spectra. Let us apply the following at-
tributes of the Welch spectra:
The signal is to be divided into 3 second segments.
The Hann (Hanning) window is to be used.
The overlap of the segments is 7/8.
The segment of the signal is to be padded with trailing zeros to smooth the spectra.
The length of trailing zeros is 100% of the segment length.
Let us compute the Welch spectra of all patient and control EEG signals. Then com-
pute mean values of each spectrum (a mean power spectral density) in the alpha
(8-12 Hz) and beta (12-25 Hz) frequency bands. After that compare the patient and
control samples of the mean power spectral density in given frequency bands.
The Mann-Whitney test indicates that there is a significant difference between the
patient and control samples (p < 0.05) in the alpha frequency band in the C4 elec-
trode. In the C3 electrode, the test indicates only a statistical tendency in the alpha
band (fig. 8).
Fig. 8. Scattering of the mean power spectral density in patients and controls. The C3 electrode
is shown at the left, the C4 electrode is shown at the right. The abscissa is a mean power spec-
tral density in the alpha frequency band in a logarithmic scale. The ordinate is a mean power
spectral density in the beta frequency band in a logarithmic scale. The patients are indicated by
diamonds and the controls are indicated by circles
Thus the Welch spectra indicate the significant difference in Parkinson's disease in the
alpha frequency band and no significant differences in the beta band. The developed
method of EEG wave packet analysis and the standard Welch spectra are complimen-
tary, that is, one can apply them simultaneously and get sufficiently different infor-
mation about the subjects.
Let us compare the developed method with the standard one based on the complex
Morlet wavelets. Firstly, compute wavelet spectrograms of all patient and control
signals in the alpha and beta frequency bands with the 0.1 Hz frequency step. Then
compute mean power spectral density of every spectrogram in every frequency band
and apply the Mann-Whitney test to compare the patient and control samples.
Information Technology and Nanotechnology (ITNT-2016) 687
Mathematical Modeling Sushkova OS, Morozov AA. et al…
The Mann-Whitney test indicates a significant difference (p < 0.03) between the pa-
tients and controls in the C3 and C4 electrodes in the alpha band (fig. 9).
Fig. 9. Scattering of the wavelet mean power spectral density in patients and controls. The
marking is the same as in the figure 8
Let us conduct another test. Compute medians of the power spectral density values of
every spectrogram in every frequency band instead of averaging the values. The val-
ues of medians are shown in the figure 10.
Fig. 10. Scattering of the medians of the wavelet power spectral density in patients and con-
trols. The marking is the same as in the figure 8
The Mann-Whitney test confirms the significant differences in the C3 and C4 elec-
trodes in the alpha band, but the probability of the first type error (p < 0.01) is better
than in the foregoing comparisons with the Welch power spectral density and the
wavelet mean power spectral density. Nevertheless, no significant differences are
revealed in the beta frequency band.
Discussion
There are contradictory results reported in the neurophysiologic papers on the Parkin-
son's disease. For instance, in [6] a significant increase of the power spectral density
in the beta frequency band in Parkinson's disease is reported. At the same time, a sig-
nificant decrease of the beta power spectral density is reported in [7,8]. Our study
shows that a detailed time-frequency dynamics of the cortex electrical activity in the
alpha and beta frequency bands is to be considered to resolve the contradiction, be-
Information Technology and Nanotechnology (ITNT-2016) 688
Mathematical Modeling Sushkova OS, Morozov AA. et al…
cause a set of sufficiently different effects present in the neighbouring frequency
bands, namely, a significant increase of the spectral density power in the alpha band
and a significant decrease of the quantity of the wave packets in the beta frequency
band. Furthermore, it is necessary to avoid an erroneous interpretation of the tails of
the alpha wave packets as an electrical activity in the beta band.
The developed method of the analysis of EEG wave packets considers separate wave
packets in the EEG signals and, therefore, analyzes the brain electrical activity more
carefully. Nevertheless, the method is complementary with respect to the standard
EEG analysis methods based on the power spectral density estimation, because these
methods may discover different significant regularities in the EEG signals.
Conclusion
The method of a brain electrical activity investigation based on the EEG wave packet
analysis is developed. The method reveals a new statistically significant effect in a
group of de novo Parkinson's disease patients. The comparison of the method with the
standard analysis methods based on the Welch spectra and complex Morlet wavelets
shows that these methods are complimentary, that is, the standard methods indicate a
significant increase of the power spectral density in the Parkinson's disease in the C3
and C4 electrodes in the alpha frequency band and the method of wave packet analy-
sis indicates a significant decrease of the quantity of the wave packets in these elec-
trodes in the nearby beta frequency range. This result demonstrates that an accurate
EEG analysis based on the investigation of the time-frequency dynamics of the elec-
trical activity is necessary to resolve the contradiction existed in the neurophysiologic
literature.
Acknowledgment
Authors are grateful to Galina D. Kuznetsova for a help and comments on the prelim-
inary versions of the paper, Yuriy V. Obukhov for a help in the statement of the prob-
lem, and Nikolay A. Kuznetsov for his critical notes on the statistical analysis meth-
od.
We acknowledge a partial financial support from the Russian Foundation for Basic
Research, grants 16-37-00426, 15-07-07846.
References
1. Lawhern V, Kerick S, Robbins K. Detecting alpha spindle events in EEG time series us-
ing adaptive autoregressive models. BMC Neuroscience, 2013; 14: 101. URL:
http://www.biomedcentral.com/1471-2202/14/101.
2. Parekh A, Selesnick IW, Rapoport DM, Ayappa I. Sleep spindle detection using time-
frequency sparsity. In IEEE Signal Processing in Medicine and Biology Symposium. Phil-
adelphia, PA: IEEE, Dec. 2014: 1-6.
Information Technology and Nanotechnology (ITNT-2016) 689
Mathematical Modeling Sushkova OS, Morozov AA. et al…
3. O’Reilly C, Nielsen T. Automatic sleep spindle detection: benchmarking with fine tem-
poral resolution using open science tools. Frontiers in Human Neuroscience, 2015; 9: 353.
DOI: 10.3389/fnhum.2015.00353.
4. Andreeva Y, Khutorskaya O. EMGs spectral analysis method for the objective diagnosis
of different clinical forms of Parkinson’s disease. J. Electromyography and Clinical Neu-
rophysiology, 1996; 36(3): 187-192.
5. Hampel FR, Ronchetti EM, Rousseeuw PJ, Stahel WA. Robust Statistics. The Approach
Based on Influence Functions. New York: John Wiley & Sons, 1986.
6. Moazami-Goudarzi M, Sarnthein J, Michels L, Moukhtieva R, Jeanmonod D. Enhanced
frontal low and high frequency power and synchronization in the resting EEG of parkin-
sonian patients. NeuroImage, 2008; 41: 985-997.
7. Pezard L, Jech R, Ruzicka E. Investigation of non-linear properties of multichannel EEG
in the early stages of Parkinson’s disease. Clinical Neurophysiology, 2001; 112: 38-45.
8. Stoffers D, Bosboom J, Deijen J, Wolters E, Berendse H, Stam C. Slowing of oscillatory
brain activity is a stable characteristic of Parkinson’s disease without dementia. Brain,
2007; 130: 1847-1860.
Information Technology and Nanotechnology (ITNT-2016) 690