Biomedical Engineering Trends in Electronics Communications and Software Part 11 ppt

Biomedical Engineering Trends in Electronics, Communications and Software

390
2.2 Approaches based on parameterization of the signal
A comparative study of various algorithms used in automatic detection methods, conducted
by Wilson and Emerson in 2002, showed that the methods use some form of
parameterization of the EEG signal usually get good results.
The first studies involving the parameterization as a tool for the detection of epileptiform
events in EEG recording were published by Gotman and Gloor (Gotman, 1976; Gotman &
Gloor, 1982) followed by the research of Webber (1994), Walckzak & Nowack (2001), Litt
(2001) and Tzallas et al. (2006) among others that have obtained promising results.
However with the advances in mathematical methods and the increasing capacity of
computer processing the investigations were directed to other approaches (Halford, 2009),
for example, the Wavelet Transform, entropy, statistical methods and/or a combination of
these and other methods (Kaneko et al. 1999; Diambra, 1999, Liu et al., 2002; Saab & Gotman,
2005; Tzallas et al., 2006; Übeyli, 2009; Kumar, 2010). Nevertheless we did not abandon the
parameterization approach (Guedes et al., 2002, Pereira 2003, Pereira et al., 2003;
Sovierzoski, 2009, Boos et al., 2010a, 2010b).
According to the literature, so far one of the most used and successful methods applied in
systems for automatic detection of paroxysms is Gotman’s (Hoef et al., 2010). This method
performs spike modeling through parameters, that in this work will be called morphological
descriptors
2
, before detection. Gotman’s method deals with the EEG signal by dividing it
into segments and sequences, both ascending and descending, which are categorized by
duration, absolute amplitude and length variation coefficient (which gives information on
the cadency of the EEG). In this system, the detection of a paroxysm occurs when the
descriptors’ values for each epoch exceeds a pre-determined threshold.
Although the literature allows access to various studies that use morphological descriptors
to characterize the EEG signal, it is necessary a detailed analysis of the applicability,

relevance and effectiveness of each descriptor that will be used.
Therefore our objective is to discuss a methodology for the preparation and evaluation of a
set of descriptors for modeling paroxysms through the use of descriptors that are already
available in the literature as well as others proposed by us in attempt to improve the
differentiation between epileptiform events and other electrographic manifestations that
occur in the signal.
3. Methodology
This section will present the recordings and methodologies used for both the development
of the descriptors’ ensemble and the experiments used as an evaluation tool for the
proposed set.
3.1 EEG recordings
All of the EEG signals used in this study belong to a database with nine records acquired
from seven adult patients with confirmed diagnosis of epilepsy. They have a sampling
frequency of 100Hz and were acquired through 24 (1 record) and 32 channels (8 records).
A bipolar montage (Fig. 2.) type zygomatic-temporal (Zygo-Db-Temp) was used, with 25
electrodes in positions Zy1, Zy2, Fp1, Fp2, F3, F4, F7, F8, F9, F10, CZ, C3, C4, T3, T4, T5, T6,

2
The use of the term morphological descriptor is because we believe that this term is more appropriate
within the context of parameters referring to morphological characteristics of a signal.
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

391
T9, T10, P3, P4, P9, P10, O1, O2 of the 10/20 system and two electrodes positioned for
acquisition of electrooculogram (EOG).
For the acquisition process the signals went through analog filtering to isolate the range of
0,5 to 40Hz. We also observed the need to perform additional filtering to remove the
baseline wandering effect (DC frequency - 0Hz) and eliminate noise caused by power line
interference (60Hz), and it was necessary to perform interpolation of the signal to a

sampling frequency of 200Hz.


(Malmivuo & Plonsey, 1995)
Fig. 2. EEG signal differences presented when a bipolar (A) and unipolar or referential (B)
montage is used. In the bipolar montage the signal is a result of potential difference between
pairs of electrodes while for the unipolar montage the signal is obtained by the difference in
potential between an electrode and a reference point (equal for the whole montage)
3.2 Morphological descriptors
The literature on the automatic detection of epileptiform events contains a considerable
amount of morphological descriptors used in different methodologies and/or developed
systems. For our experiments we selected the descriptors most reported in literature: the
maximum amplitude of the event, event duration, the length variation coefficient, crest
factor and entropy.
The maximum amplitude and duration of the event are self-explanatory. The length
variation coefficient – used to measure the regularity of the signal – is the ratio of standard
deviation and the mean value of the signal. The crest factor is the difference between the
maximum and minimum amplitudes, divided by the standard deviation (Webber et al.,
1994). The entropy, reported in several studies – e.g. Quiroga (1998), Esteller (2000),
Srinivasan et al. (2007) and Naghsh-Nilchi & Aghashahi (2010) - provides a value for the
complexity of the signal under analysis.
These descriptors are widely used, however they may not guarantee the complete
differentiation between the events presented by the recordings and also because of this the
existing systems for automatic detection have only a moderate performance. Thus, through
a detailed analysis of the EEG signals that are being used, new descriptors based on the
physical and/or morphological signal can be developed in attempt to improve the
performance of the automatic detection process.
The main focus for the development of new descriptors was to find characteristics in the EEG
signals that further highlighted the epileptiform events from other types of events. The latter
are called non-epileptiform events (Fig. 3.) and for our database they are represented by:

Biomedical Engineering Trends in Electronics, Communications and Software

392
a. normal background EEG activity;
b. alpha waves;
c. blinks;
d. artifacts originated from EMG (muscle activity), external electromagnetic interference,
among others.

0 10 20 30 40 50 60 70 80 90 100
-200
-150
-100
-50
0
50
100
Amplitude (
μ
V)
( a )
0 10 20 30 40 50 60 70 80 90 100
-200
-150
-100
-50
0
50
100
( b )

0 10 20 30 40 50 60 70 80 90 100
-200
-150
-100
-50
0
50
100
Tim e (10
-2
s)
Amplitude (
μ
V)
( c )
0 10 20 30 40 50 60 70 80 90 100
-200
-150
-100
-50
0
50
100
Tim e (10
-2
s)
( d )

Fig. 3. Morphology of the main non-epileptiform events found in our EEG signals database

0 10 20 30 40 50 60 70 80 90 100
-80
-40
0
40
80
120
Amplitude (
μ
V)
0 10 20 30 40 50 60 70 80 90 100
-60
-40
-20
0
20
40
60
0 10 20 30 40 50 60 70 80 90 100
-80
-60
-40
-20
0
20
40
Time (10
-2
s)
Amplitude (

μ
V)
0 10 20 30 40 50 60 70 80 90 100
-100
-50
0
50
100
Tim e (10
-2
s)

Fig. 4. Morphology presented by the epileptiform events in the recordings under analysis
Looking at the obtained records we realized that due to the use of a bipolar montage (Fig.2)
the epileptiform events can appear in four different ways (Fig. 4.). In other words, because
of the type of montage the spikes and sharp waves may appear with both electronegative
and electropositives amplitude peaks, however to be considered a paroxysm they still have
to be followed by a slow wave.
The basic morphological characteristics of an epileptic event are related to their amplitude
and duration. The spikes have duration of 20 to 70ms, while a sharp wave has duration of 70
to 200ms. Since both events can be a paroxysm and making a distinction between them
makes little sense from a clinical point of view, we can consider that the duration
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

393
epileptiform events varies from 20 to 200ms. The amplitudes values of both spikes and
sharp waves are also varied but when considering them epileptiform events the amplitude
(module value) usually lies between 20μV and 200μV (Niedermeyer, 2005). Examples of
morphological descriptors related to the amplitude and duration of a typical epileptiform

event are (Fig. 5.):
• maximum amplitude (Amax);
• minimum amplitude (Bmin);
• difference between the points of occurrence of extreme amplitude (Tdif);
• difference between the maximum and minimum amplitudes (DifAB).

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
-150
-140
-130
-120
-110
-100
-90
-80
-70
-60
-50
-40
-30
-20
-10
0
10
20
30
40
50
60
70

Time (s)
Ampitude (
μ
V)
Amax
Bmin
DifAB
Tdif

Fig. 5. Morphological descriptors related to the amplitude and duration of paroxysms

0 10 20 30 40 50 60 70 80 90 100
-200
-100
0
100
200
Ampitude (
μ
V)
maximum (event)
minimum (event)
Tim e ( 10
-2
s )
minimum (epoch)

Fig. 6. EEG signal presenting maximum amplitude corresponding to an epileptiform event
and minimum amplitude corresponding to another (different) event
Also regarding the amplitudes within the epoch under review (in this case 1 second of the

signal) the points of maximum and minimum amplitude may not belong to the same event
(Fig. 6.). Analyzing this fact, we could see that to be a paroxysm (event we want to correctly
identify) the event should have a time difference between maximum and minimum
amplitudes in the range of 35 to 100ms (half duration the slowest event). For this, as
illustrated in Fig 7, we determined a 300ms segment centered at the event appearing in the
epoch under review and within this segment we calculated the following descriptors:
maximum amplitude (Amax_pts); minimum amplitude (Bmin_pts); distance between
extreme amplitudes (DifAB_pts) and time difference (Tdif_pts) between the maximum and
minimum amplitudes.
Another feature that can be observed is that an epileptiform event, particularly the spike,
has more acute peaks when compared to the obtuse peaks of alpha waves or blinks (Fif. 3b
and Fig. 3c). This fact allows another opportunity to discriminate between events since the

Biomedical Engineering Trends in Electronics, Communications and Software

394
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
-150
-140
-130
-120
-110
-100
-90
-80
-70
-60
-50
-40
-30

-20
-10
0
10
20
30
40
50
60
70
Time (s )
Ampitude (
μ
V)
Amax_pts
Bmin_pts
DifAB_pts
Tdif_pts

Fig. 7. Maximum amplitude (Amax_pts), minimum amplitude (Bmin_pts), distance between
extreme amplitudes (Tdif_pts) and time difference between amplitudes (DifAB_pts), all
within the 300ms segment centered on the event under analysis
process of automatic detection can confused them, which is a detrimental factor to the
system performance. Based on these observations we analyzed the vertex angle of the peaks
through the extreme amplitudes and zero crossing points adjacent to the beginning and the
end of the event.

0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75
-150
-140

-130
-120
-110
-100
-90
-80
-70
-60
-50
-40
-30
-20
-10
0
10
20
30
40
50
60
70
Ampitude (
μ
V)
dneg
dpos
Time (s )
θ
n
trp

trn
θ
p

Fig. 8. Vertex angle of positive and negative epileptiform event, calculated from the
maximum and minimum amplitude, respectively
The calculated angles (Fig. 8.), taking an epileptiform event as example, refer to the angle
influenced by the peak’s initial inclination and the angle that suffers influence of beginning
slope of the slow wave. Based on the calculation of these angles (θp and θn) we determined
other descriptors:
• base of the peaks directly adjacent to the beginning and the end of the event (dpos and
dneg, depending in order of appearance of the peaks);
• angle of the analyzed event apex (θ);
• tangents of the angles of peak apex (tgp and tgn);
• tilt of the slopes directly adjacent to the beginning and the end of the event (trp and
trn);
• event basis (dbase).
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

395
The morphology of a paroxysm can also often be confused with the morphology of artifacts
(from various sources) present in the EEG signal. However, as can be seen in Fig. 9. the
typical waveforms of these noises usually have a relative high frequency. This means that
the high amplitudes appear with minimum time differences between them, which are the
opposite of paroxysms that usually have more widely spaced peaks because they are always
followed by a slow wave.

0 10 20 30 40 50 60 70 80 90 100
-120

-100
-80
-60
-40
-20
0
20
40
60
80
Tim e ( x 10
-2
s )
Ampitude (
μ
V)

Fig. 9. Example of typical morphology of an artifact (noise) present in the EEG signal

0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00
-150
-140
-130
-120
-110
-100
-90
-80
-70
-60

-50
-40
-30
-20
-10
0
10
20
30
40
50
60
70
80
90
Time (s )
Ampitude (
μ
V)
Amax_i
Bmin_i
Bmin_f
tA_i tA_f
tB_i tB_f
Amax_f
Amax
Bmin
Final RegionInitial Region

Fig. 10. Descriptors for the differentiation between epileptiform events and artifacts,

considering distances (time) between the points of maximum and minimum amplitude
The descriptors proposed to make the distinction between noise and epileptiform events can
be based on relations of time and amplitude differences in the epoch when dividing it in two
regions (initial and final) adjacent to the event. Experiments were performed and from them
we projected the following descriptors:
• amplitude and time difference between maximum amplitudes of the event (Amax),
initial (Amax_i) and final regions (Amax_f): DifA_i, tA_i, DifA_f and tA_f;
• amplitude and time difference between minimum amplitudes of the event (Bmin),
initial (Bmin_i) and final regions (Bmin_f): DifB_i, tB_i, DifB_f, tB_f.
Further analysis of the morphology and other characteristics of events that occur in the EEG
recordings can be performed. In this research it is proposed only the addition of descriptors
based on the classical statistical indices of average, standard deviation and variance. These
Biomedical Engineering Trends in Electronics, Communications and Software

396
descriptors, were calculated for both the epoch under analysis (one second) and the 300ms
segment. Thus, considering the descriptors selected from the literature and those we
developed after a review of the recordings, we obtained a final set of 45 morphological
descriptors (Table 1).

Origin if the descriptors Descriptors identifications
Amplitude Amax, Bmin, DifAB, Amax_pts, Bmin_pts, DifAB_pts
Duration Tdif, Tdif_pts, T
Vertex angle of the peaks θ, θp, θn, dbase, dpos, dneg, trp, trn, tgp, tgn
Initial region of the epoch Amax_i, Bmin_i, DifA_i, tA_i, DifB_i, tB_i
Final region of the epoch Amax_f, Bmin_f, DifA_f, tA_f, DifB_f, tB_f
Statistical indexes
1

desvio, media, var, coef, CF, desvioC, mediaC, varC, coefC,

CFC
Entropy
1,2
entrop_log, entrop_norm, entrop_logC, entrop_normC
1
The letter ‘C’ at the end of the identification means that the descriptor was calculated for the
segment of 300ms.
2
We calculated two types of entropy: normalized (norm) and logarithm of "energy".
Table 1. Summary of the elements that compose the final set of 45 morphological descriptors
selected and developed for this research
3.3 Morphological descriptors evaluation
In the previous item (3.2) 45 morphological descriptors were presented. Some of them were
chosen among those universally used and others were defined in our previous work.
After the creation of the descriptors’ set it is necessary to analyze this ensemble in order to
verify the significance of each element of the group in the differentiation of events. For this
research we chose to use correlation analysis and application of Hotelling’s T² test (Härdle &
Simar, 2007) for individual assessment and Artificial Neural Networks (Eberhart & Dobbins,
1990; Zurada, 1992; Haykin, 1994 ) to verify the complete set performance.
The correlation analysis was made evaluating the correlation matrices of descriptors for
pairs of events. We examined the correlation between morphological descriptors calculated
from epochs containing paroxysm and epochs with non-epileptiform (blinks, artifacts, alpha
waves and background EEG activity). The criterion for possible exclusion of any element
(descriptor) of the designed set was the existence of high correlation values (above 50%) for
all pairs of events considered.
The Hotelling’s T² test consisted in calculating the difference between the values of each
descriptor in epochs with epileptiform transients and epochs with non-epileptiform events.
The assessment of this test was made comparing the results of these differences with a
predetermined T² critical value (a threshold). Based on this test a descriptor is considered
relevant when its T² result is greater than the pre-determined critical value.

Some descriptors such as the tangents of the positive (θp) and negative (θn) angles, length
variation coefficient (coef) and crest factor (CF) had T² test result relatively close to the
critical value and thus these elements could have been removed from the set. However as
the correlation value achieved by these same descriptors was not high and their exclusion
did not affect significantly the sensitivity and specificity of the neural networks
implemented in this study. We chose to not exclude them from the final set.
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

397
For the verification that the descriptors can indeed provide sufficient information so a
classifier can make the discrimination between events the set was arranged at the input of
several Artificial Neural Networks.
The networks used are all Feedforward Multilayer Perceptron with Backpropagation
algorithm and supervised learning. The basic architecture of each network was of an input
layer with 45 neurons and output layer with only one neuron. The number of neurons in the
hidden layer and the application of input stimuli normalization
3
were varied in each of the
networks so we could find the best configuration and analyze the effect of this
normalization. Some other features of neural networks implemented are:
• activation function of output and hidden layers: hyperbolic tangent;
• number of neurons in the hidden layer (N): 7 to 11 neurons;
• batch update of the synaptic weights (after every training epoch);
• learning rate and momentum were respectively: 0,01 and 0,9.
Finally, the training and test of networks were made with two different compositions of files
(Table. 2.): a set of files classified only by the presence or absence of paroxysms and another
set where the files were classified by type of event (sharp waves, spikes, blinks, normal
background EEG activity, alpha waves and artifacts).

Composition Process Signal classification used
No. of
files
Epileptiform event 47
Training
Non-epileptiform event 73
Epileptiform event 30
Composition I
Test
Non-epileptiform event 23
Sharp wave 10
Epileptiform event
Spike 10
EEG background
activity
5
Alpha waves 10
Artifacts 5
Composition II
Training
and testª
Non-epileptiform event
Blinks 5
Table 2. Composition of files according to different classifications of EEG signals events,
used for training and tests of the neural networks created
4. Results
Several networks with the same basic architecture and features showed in the previous
section were trained and tested using both types of file composition (Table 2.). The
normalization of input stimuli was tested in all implemented networks.
The set of descriptors (computed for each file) were attributed directly to the networks’

input and the stopping criteria for training, used in our experiments, was the minimum
error (1%) and the maximum number of iterations allowed (100.000 epochs).

3
The term normalization refers to the operation of correcting the amplitude of EEG recordings in which
the maximum amplitude is greater than the one of a paroxysm (± 200μV). The applied correction is the
ratio between the signal and its mean value.
Biomedical Engineering Trends in Electronics, Communications and Software

398
The best results obtained after the simulations with all these networks are presented in
Table 3, where the following statistical indices can be observed:
• Success rate (SR);
• True positive (TP), true negative (TN), false positive (FP) and false negative (FN);
• Sensibility (SE) e specificity (SP);
• Positive predictive value (PPV) and negative predictive value (NPV).

ANN specifications SR TP TN FP FN SE SP PPV NPV
8N hidden / 10
5
epochs
a
81% 27 19 4 3 0,90 0,83 0,87 0,86
9N hidden / 10
5
epochs
a
79% 27 20 3 3 0,90 0,87 0,90 0,87
8N hidden / 10
5

epochs
a,c
79% 27 16 7 3 0,90 0,70 0,79 0,84
8N hidden / 10
5
epochs
b
80% 18 24 1 2 0,90 0,96 0,95 0,92
9N hidden / 11863 epochs
b
89% 17 24 1 3 0,85 0,96 0,94 0,89
9N hidden / 12026 epochs
b,c
68% 19 19 6 1 0,95 0,76 0,76 0,95
a Training and test with files from composition I.
b Training and test with files from composition
II.
c The input stimuli was normalized.
Table 3. Best results achieved with the Artificial Neural Networks created
According to results presented in Table 3 the use of files with signals classified by the
occurrence of paroxysms showed success rate (the correct identification of test signals) of
79% whereas with the files of the composition II this rate was around 90%. The best network
implementations for each type of files showed sensitivity of 90% and 85% and specificity of
87% and 96%.
The effect of normalizing the network’s input stimuli that we observed during the
simulations was a reduction in the specificity values due to the number of false positives
generated (for example, for the network with nine hidden neurons the false positives
increased from one to six).
5. Conclusions
The use and determination of morphological descriptors seems to be simple because it is a

direct data collection with relatively basic calculations such as, for example, calculating the
dimensions of amplitude and duration of the event. However, this process requires a priori
knowledge of information about the system or entity which characteristics will be cataloged. In
other words, for the case of automatic detection of epileptiform events in EEG recordings is
necessary to carry out preliminary studies about the morphology of the signals to be analyzed.
Another significant aspect when using morphological descriptors is the assessment of the
selected descriptors as input of the classifier used. It is important to perform an evaluation
to demonstrate the contribution of each descriptor for the capability of the ensemble in
making the distinction between events of interest. In this study we used correlation analysis
and Hotelling’s T² test to identify which descriptors could be excluded from the created set
in order to provide a performance improvement of the automatic detection process. The
methods applied for this assessment did not result in significantly high improvements in the
automatic detection, but this does not invalidate its use because the classifier (neural
network) used on the experiments showed promising results.
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

399
Thus, it becomes necessary to study other advanced and robust analysis tools that can
within a tolerance (error) threshold, provide more consistent results. Therefore we are using
multivariate analysis (Principal Component Analysis, Independent Component Analysis)
alone or in combination with other statistical techniques for assessing the relevance of the
descriptors in attempt to optimize the size of the set needed to perform automatic detection
through neural networks (or other classifier) without causing significant performance loss
for the system in which the descriptors are inserted.
6. References
Abibullaev B.; Seo H. D. & Kang W. (2009a) A Wavelet Based Method for Detecting and
Localizing Epileptic Neural Spikes in EEG. Proceedings of the 2nd International
Conference on Interaction Sciences: Information Technology, Culture and Human. pp.
702-707, 9781605587103, Seoul, Korea, November 2009, ACM, New York.

Abibullaev B.; Kim M. S. & Seo H. D. (2009b) Seizure Detection in Temporal Lobe Epileptic
EEGs Using the Best Basis Wavelet Functions. Jornal of Medical Systems. Vol.34,
No.4, May 2009, pp. 755-765, 0148-5598.
Adeli, H.; Zhou, Z. & Dadmehr, N. (2003). Analysis of EEG records in an epileptic patient
using wavelet transform. Jornal of Neuroscience Methods. Vol.123, February 2003, pp.
69-87, 0165-0270.
Argoud, F.I.M.; De Azevedo, F.M.; Marino Neto, J. & Grillo, E. (2006). SADE3: an effective
system for automated detection of epileptiform events in long-term EEG based on
context information. Medical & Biological Engineering & Computing. Vol.44, No.6,
June 2006, Springer, pp. 459–470, 0140-0118.
Boos, C. F.; Pereira, M. C. V.; Argoud, F. I. M.; Azevedo, F. M (2010a). Analysis and
definition of morphological descriptors for automatic detection of epileptiform
events in EEG signals with artificial neural networks. Proceedings of the 3rd IEEE
International Conference on Computer Science and Information Technology. Vol.5, pp.
349-353, 978-1-4244-5537-9, Chengdu, Sichuan, China, July 2010, IEEE Press.
Boos, C. F.; Pereira, M. C. V.; Argoud, F. I. M.; Azevedo, F. M (2010b). Morphological
descriptors for automatic detection of epileptiform events. Proceedings of the 32nd
Annual International Conference of the IEEE Engineering in Medicine and Biology
Society. pp. 2435-2438, 978-1-4244-4124-2, Buenos Aires, Argentina, August-
September 2010, IEEE Press.
Coimbra, A. J. F.; Marino J.; Freitas, C. G.; Azevedo, F. M., Barreto, J. M. (1994). Automatic
Detection of Sleep-Waking States Using Kohonen Neural Networks. Proceedings of
the I Congresso Brasileiro de Redes Neurais, Itajubá, Brazil, October 1994, pp. 327-331.
Diambra, L.; Fiqueiredo, J.C.B.; Malta,C.P. (1999). Epileptic Activity Recognition in EEG
Recording, Physica A, Vol.273, No.3, November 1999, Elsevier, pp.495-505, 0378-
4371.
Eberhart, R & Dobbins, R. (1990). Neural Network PC Tools: A Practical Guide. Academic Press,
0122286405, San Diego, California.
Esteller, R. (2000). Detection of Seizure Onset in Epileptic Patients from Intracranial EEG
Signal. PhD. Thesis, School of Electrical and Computer Engineering Georgia

Institute of Technology, 2000.
Fisher, R.S.; Boas, W.E.; Blume,W., Elger, C.; Genton, P.; Lee, P.; Engel Jr, J. (2005) Epileptic
Siezures and Epilepsy: Definitions Proposed by the International League Against
Biomedical Engineering Trends in Electronics, Communications and Software

400
Epilepsy (ILAE) and International Bureau for Epilepsy (IBE). Epilepsia, Vol.46, No.4,
March 2005, p.470-472, 1528-1167.
Gotman, J. (1982). Automatic recognition of epileptic seizures in the EEG,
Electroencephalography and Clinical Neurophysiology, Vol.54, No.5, November 1982,
pp. 530-540, 0013-4694.
Gotman, J.; Gloor, P. (1976). Automatic recognition and quantification of interictal epileptic
activity in the human scalp EEG, Electroencephalography and Clinical Neurophysiology,
V.41, No.5, November 1976, pp. 513-529, 0013-4694.
Guedes, J. R. ; Pereira, M. C. ; Azevedo, F. M. ou de Azevedo, F.M. (2002). Parameterization
of the EEG Signal Applied to the Detection of Epileptiform Events. Proceedings of the
2nd European Medical and Biological Engineering Conference, Vol.3(1), pp. 444-445,
Vienna, Austria, December 2002, IFMBE.
Halford, J. J. (2009). Computerized epileptiform transient detection in the scalp
electroencephalogram: Obstacles to progress and the example of computerized
ECG interpretation. Clinical Neurophysiology, Vol.120, No.11, November 2009, pp.
1909–1915, 1388-2457.
Haykin, S. (1994) Neural Networks: A Comprehensive Foundation, Macmilliam College
Publishing Company, 0-13-226556-7, Englewood Cliffs.
Hoef, L.; Elgavish, R.; Knowlton, R. C. (2010). Effect of Detection Parameters on Automated
Electroencephalography Spike Detection Sensitivity and False-Positive Rate. Journal
of Clinical Neurophysiology, Vol.27, No.1, February 2010, pp. 12-16, 0736-0258
Hoffmann, K.; Feucht, M.; Witte, H.; Benninger, F & Bolten, J. (1996). Analysis and
classification of interictal spikes discharges in Benign Partial Epilepsy of Childhood
on the basis of the Hilbert transform, Neuroscience Letters, Vol. 211, No.3, June 1996,

pp. 195-198, 0304-3940.
Indiradevi, K.P.; Elias, E.; Sathidevi, P.S.; Nayak, S.D. & Radhakrishnan, K. (2008). A multi-
level wavelet approach for automatic detection of epileptic spikes in the
electroencephalogram, Computers in Biology and Medicine, V.38, No.7, July 2008, pp.
805-816, 0010-4825.
Kalayci, T. & Özdamar, O. (1995). Wavelet preprocessing for automated neural network
detection of EEG spikes, IEEE Engineering in Medicine and Biology Magazine, Vol.14,
No.2, March 1995, pp. 160-166, 0739-5175.
Kaneko, H.; Suzuki, S. S.; Akamatsu, M. (1999). Multineuronal Spike Classification Based on
Multisite Electrode Recording, Whole-Waveform Analysis, and Hierarchical
Clustering. IEEE Transactions on Biomedical Engineering, Vol.46, No.3, March 1999,
pp. 280-290, 0018-9294.
Khan, Y.U. & Gotman, J. (2003). Wavelet based automatic seizure detection in intracerebral
electroencephalogram, Clinical Neurophysiology, Vol.114, No.5, May 2003, pp. 898-
908, 1388-2457.
Kim, K. H.; Kim, S. J. (2000). Neural Spike Sorting Under Nearly 0-dB Signal-to-Noise Ratio
Using Nonlinear Energy Operator and Artificial Neural-Network Classifier. IEEE
Transactions on Biomedical Engineering, Vol.47, No.10, October 2000, pp. 1406-1411,
1406-1411.
Kumar, S. P.; Sriraam, N.; Benakop, P.G.; Jinaga, B.C. (2010). Entropies based detection of
epileptic seizures with artificial neural network classifiers, Expert Systems with
Applications, Vol.37, No.4, April 2010, pp. 3284-3291, 0957-4174.
Automatic Detection of Paroxysms in EEG Signals using Morphological Descriptors
and Artificial Neural Networks

401
Litt, B.; Esteller, R; Echauz, J; D'Alessandro, M.; Shor, R.; Henry, T.; Pennell, P.; Epstein, C.;
Bakay, R.; Dichter, M. & Vachtsevanos, G. (2001). Epileptic Seizures May Begin
Hours in Advance of Clinical Onset: A Report of Five Patients, Neuron, Vol.30,
No.1, April 2001, pp. 51-64, 0896-6273.

Liu, H.S.; Zhang, T. & Yang, F.S. (2002). A multistage, multimethod approach for automatic
detection and classification of epileptiform EEG, IEEE Transactions on Biomedical
Engineering, Vol. 49, No.12, December 2002, pp. 1557-1566, 0018-9294.
Malmivuo, J.; Plonsey, R. (1995). Bioelectromagnetism: Principles and Applications of Bioelectric
and Biomagnetic Fields, Oxford University Press, 0-19-50-5823-2, New York.
Mohamed, N.; Rubin, D.M. & Marwala, T. (2006). Detection of epileptiform activity in
human EEG signals using Bayesian neural networks, Neural Information Processing –
Letters and Reviews, Vol.10, No.1, January 2006, pp. 231–237, 1738-2572.
Naghsh-Nilchi, A.R. & Aghashahi, M. (2010). Epilepsy seizure detection using eigen-system
spectral estimation and Multiple Layer Perceptron neural network, Biomedical
Signal Processing and Control, Vol.5, No.2, April 2010, pp. 147-157, 1746-8094.
Niedermeyer, E. & Silva F. L. (1993). Electroencephalography: Basic Principles, Clinical
Applications, and Related Fields. Williams & Wilkins, 0-683-06511-4, Philadelphia.
Ocak, H. (2008). Automatic detection of epileptic seizures in EEG using discrete wavelet
transform and approximate entropy, Expert Systems with Applications, Vol.36, No.2,
Part 1, March 2009, pp. 2027-2036, 0957-4174.
Ocak, H. (2008). Optimal classification of epileptic seizures in EEG using wavelet analysis
and genetic algorithm, Signal Processing, Vol.88, No.7, July 2008, pp. 1858-1867,
0165-1684.
Oweiss, K.G. & Anderson, D.J. (2001). Noise reduction in multichannel neural recordings
using a new array wavelet denoising algorithm, Neurocomputing,Vol.38-40, June
2001, pp. 1687-1693, 0925-2312.
Pang, C.C.C.; Upton, A.R.M.; Shine, G. & Kamath, M.V. (2003). A Comparison of Algorithms
for Detection of Spikes in the Electroencephalogram, IEEE Transactions on Biomedical
Engineering, Vol.50, No.4, April 2003. pp. 521-526, 0018-9294.
Pereira, M. C. V. (2003). Avaliação de técnicas de pré-processamento de sinais do EEG para detecção
de eventos epileptogênicos utilizando redes neurais artificais. Thesis (PhD), Biomedical
Engineering Institute, University Federal of Santa Catarina, 2003.
Pereira, M. C. V.; Azevedo, F.M. & Argoud, F. I. M. (2003). Investigation About Pre-
Processing in the Input of an Artificial Neural Network for Analysis of

Epileptogenic Events in EEG Signals. Proceedings of the World Congress on Medical
Physics and Biomedical Engineering, pp.103-103, Sydney, Australia, IFMBE.
Pillay, J. & Spearling, M. R. (2006). Interictal EEG and the Diagnosis of Epilepsy, Epilepsia,
Vol.47, No.1, October 2006, pp. 14-22, 1528-1167.
Quiroga, R. Q. (1998). Quantitative Analysis of EEG Signals: Time-Frequency Methods and Chaos
Theory. Thesis (PhD), Institute of Physiology, Medical University Lübeck and
Institute of Signal Processing, Medical University Lübeck, 1998
Quiroga, R.Q.; Sakowitz, O.W.; Basar, E. & Schürmann, M. (2001). Wavelet Transform in the
analysis of the frequency composition of evoked potentials, Brain Research Protocols,
Vol.8, No.1, August 2001, pp. 16-24, 1385-299X.
Saab. M. E. & Gotman, J. (2005). A system to detect the onset of epileptic seizures in scalp
EEG, Clinical Neurophysiology, Vol.116, No.2, February 2005, pp. 427-442, 1388-2457.
Biomedical Engineering Trends in Electronics, Communications and Software

402
Sanei, S. & Chambers, J.A. (2007). EEG Signal Processing, John Wiley & Sons, 978-0-470-
02581-9, West Sussex.
Scolaro, G.R. & Azevedo, F.M. (2010). Classification of Epileptiform Events in Raw EEG
Signals using Neural Classifier, Proceedings of the 3rd IEEE International Conference on
Computer Science and Information Technology. Vol.5, pp. 368-372, 978-1-4244-5537-9,
Chengdu, Sichuan, China, July 2010, IEEE Press.
Sovierzoski, M. A. (2009). Avaliação de descritores morfológicos na identificação de eventos
epileptiformes. Thesis (PhD), Biomedical Engineering Institute, University Federal of
Santa Catarina, 2009.
Srinivasan, V.; Eswaran, C. & Sriraam, N. (2007). Approximate Entropy-Based Epileptic EEG
Detection Using Artificial Neural Networks, IEEE Transactions on Information
Technology in Biomedicine, Vol.11, No.3, May 2007, pp.288-295, 1089-7771.
Subasi, A. (2007). Application of adaptive neuro-fuzzy inference system for epileptic seizure
detection using wavelet feature extraction, Computers in Biology and Medicine,
Vol.37, No.2, February 2007, pp. 227-244, 0010-4825.

Tzallas, A.T.; Karvelis, P. S.; Katsis, C. D.; Fotiadis, D.I., Giannopoulus, S. & Konitsiotis, S.
(2006). A Method for Classification of Transient Events in EEG Recordings:
Application to Epilepsy Diagnosis, Methods of Information in Medicine, Vol.45, No.6,
March 2006, pp. 610-621, 0026-1270.
Übeyli, E. D. (2009). Statistics over features: EEG signals analysis, Computers in Biology and
Medicine, Vol.39, No.8, August 2009, pp. 733-741, 0010-4825.
Walczak, S. & Nowack, W. J. (2001) An Artificial Neural Network Approach to Diagnosing
Epilepsy Using Lateralized Bursts of Theta EEGs, Journal of Medical Systems, Vol.25,
No.1, February 2001, pp. 1-22, 0148-5598.
Webber, W. R. S.; Litt, B.; Wilson, K. & Lesser, R. P. (1994). Practical Detection of
Epileptiform Discharges (EDs) in the EEG Using an Artificial Neural Network: a
Comparison of Raw and Parameterized EEG Data, Electroencephalography and
Clinical Neurophysiology, Vol.91, No.3, September 1994, pp. 194-204, 0013-4694.
Wilson, S. B.; Emerson, R. (2002). Spike Detection: a Review and Comparison of Algorithms,
Clinical Neurophysiology, Vol.113, No.12, December 2002, pp. 1873-1881, 1388-2457.
Wilson, S.B.; Scheuerb, M.L.; Emerson R.G. & Gabor, A.J. (2004). Seizure detection:
evaluation of the Reveal algorithm, Clinical Neurophysiology, Vol.115, No.10,
October 2004, pp. 2280-2291, 1388-2457.
Zurada, J. M. (1992). Introduction to Artificial Neural Systems, West Publishing Company, 0-
314-93391-3, St. Paul.


21
Multivariate Frequency Domain
Analysis of Causal Interactions in
Physiological Time Series
Luca Faes and Giandomenico Nollo
Department of Physics and BIOtech Center
University of Trento
Italy

1. Introduction
A common way of obtaining information about a physiological system is to measure one or
more signals from the system, consider their temporal evolution in the form of numerical
time series, and obtain quantitative indexes through the application of time series analysis
techniques. While historical approaches to time series analysis were addressed to the study
of single signals, recent advances have made it possible to study collectively the behavior of
several signals measured simultaneously from the considered system. In fact, multivariate
(MV) time series analysis is nowadays extensively used to characterize interdependencies
among multiple signals collected from dynamical physiological systems. Applications of
this approach are ubiquitous, for instance, in neurophysiology and cardiovascular
physiology (see, e.g., (Pereda et al., 2005) and (Porta et al., 2009) and references therein). In
neurophysiology, the time series to be analyzed are obtained, for example, sampling
electroencephalographic (EEG) or magnetoencephalographic (MEG) signals which measure
the temporal dynamics of the electro-magnetic fields of the brain as reflected at different
locations of the scalp. In cardiovascular physiology, the time series are commonly
constructed measuring at each cardiac beat cardiovascular and cardiorespiratory variables
such as the heart period, the systolic/diastolic arterial pressure, and the respiratory flow. It
is well recognized that the application of MV analysis to these physiological time series may
provide unique information about the coupling mechanisms underlying brain dynamics
and cardiovascular control, and may also lead to the definition of quantitative indexes
useful in medical settings to assess the degree of mechanism impairment in pathological
conditions.
MV time series analysis is not only important to detect coupling, i.e., the presence or absence
of interactions, between the considered time series, but also to identify driver-response
relationships between them. This problem is a special case of the general question of
assessing causality, or cause-effect relations, between (sub)systems, processes or phenomena.
The assessment of coupling and causality in MV processes is often performed by linear time
series analysis approaches, i.e. approaches in which a linear model is supposed to underlie
the generation of temporal dynamics and interactions of the considered signals (Kay, 1988;
Gourevitch et al., 2006). While non-linear methods are continuously under development

Biomedical Engineering Trends in Electronics, Communications and Software

404
(Pereda et al., 2005; Faes et al., 2008), the traditional linear approach remains of great interest
for the study of physiological signals, mainly because it has the important advantage to be
strictly connected to the frequency-domain representation of multichannel data. Indeed,
physiological signals such as the brain and cardiovascular ones are rich of oscillatory
content and thus lend themselves to spectral representation. Typical examples of
physiological rhythms are the EEG dynamics, typically observed within the well-bounded
frequency bands from delta to gamma (Nunez, 1995), and the cardiovascular oscillations,
characterized by spectral peaks within the so-called low frequency (LF, ~0.1 Hz) and high
frequency (HF, synchronous with respiratory activity) bands (Akselrod et al., 1981). As a
consequence, the linear frequency-domain evaluation of coupling and causality constitutes
an eligible approach to characterize the interdependence among specific oscillations
manifested within the same frequency band in two or more physiological signals.
While an unique and universally accepted definition of causality does not exist, in time
series analysis inference about cause-effect relationships is commonly based on the notion
introduced by Nobel Prize winning Clive Granger (Granger, 1969). Granger causality was
mathematically formalized within a linear time-domain framework widely applied in
economy and finance but rapidly spread to other fields including the analysis of
physiological time series. This notion of causality is defined in terms of predictability and
exploits the direction of the flow of time to achieve a causal ordering of dependent
processes. The definition may be contextualized in a different way for bivariate (based on
two signals only) and MV (based on more than two signals) analysis; in the MV formulation,
a distinction between direct causality from one series to another and indirect causality (i.e.,
causality between two series mediated by other series) is achieved (Faes et al., 2010b).
Moreover, while the most intuitive definition of causality accounts for lagged effects only
(i.e., effects of the past of a time series on the present of another), the concept of
instantaneous causality, describing influences which occur within the same lag, is crucial for
the evaluation of causal relationships among processes (Lutkepohl, 1993). Finally, the

different facets of the concept of causality may be related to the concept of coupling between
two processes, according to which the presence or absence of an interaction is detected and
measured, but the directionality of such interaction is not elicited.
The notions of causality and coupling are commonly formalized in the context of a MV
autoregressive (MVAR) representation of the available time series, which allows to derive
time- and frequency-domain pictures of these concepts respectively through the model
coefficients and through their spectral representation. Accordingly, several frequency domain
measures of causality and coupling have been introduced and applied in recent years.
Coupling is traditionally investigated by means of the coherence (Coh) and the partial
coherence (PCoh), classically known, e.g., from Kay (1988) or (Bendat & Piersol, 1986).
Measures able to quantify causality in the frequency domain have been proposed more
recently: the most used are the directed transfer function (DTF) (Kaminski & Blinowska, 1991),
the directed coherence (DC) (Baccala et al., 1998), and the partial directed coherence (PDC)
(Baccala & Sameshima, 2001). All these measures have been used extensively for the analysis
of physiological time series, and applications showing their usefulness for the interpretation of
interaction mechanisms among, e.g., EEG rhythms or cardiovascular oscillations, are plentiful
in the literature (see, for instance, (Porta et al., 2002; Schlogl & Supp, 2006; Astolfi et al., 2007;
Faes & Nollo, 2010a)). Despite this, several issues have to be taken into account for their correct
utilization. While the relationships existing among these indices are generally understood, and
most of the properties linking these measures to the different concepts of causality and
Multivariate Frequency Domain Analysis of Causal Interactions in Physiological Time Series

405
coupling are known, an organic joint description and contextualization in relation to the
underlying time domain concepts is lacking. Also for this reason, the interpretation of
frequency-domain coupling and causality measures is not always straightforward, and this
may lead to an erroneous description of connectivity and related mechanisms. Examples of
ambiguities emerged in the interpretation of these measures are the debates about the ability
of PCoh to measure some forms of causality (Albo et al., 2004; Baccala & Sameshima, 2006),
and about the specific kind of causality which is reflected by the DTF and DC measures

(Kaminski et al., 2001; Baccala & Sameshima, 2001; Eichler, 2006). An aspect which is perhaps
more problematic regards the structure of the model used to represent the data prior to
computation of the frequency domain measures, which commonly accounts for lagged but not
for instantaneous effects among the series. Despite this, the significance of instantaneous
correlations among the series is almost never tested in practical applications, and the possible
effects on coupling and causality measures of forsaking such correlations have not been
investigated thoroughly. Very recent studies have suggested that neglecting instantaneous
interactions in the model representation may lead to heavily modified connectivity patterns
(Hyvarinen et al., 2008; Faes & Nollo, 2010b).
The mission of this chapter is to enhance the theoretical interpretability of the available
frequency domain measures of coupling and causality derived from the MVAR
representation of multiple time series. To this end, a common framework for the definition
of Coh, PCoh, DC/DTF, and PDC is provided on the basis of the frequency domain MVAR
representation, and is exploited to relate the various measures to each other as well as to the
specific coupling or causality definitions which they underlie. The chapter is structured as
follows: Sect. 2 presents a comprehensive definition of the various forms of causality and
coupling that can be observed in MV processes; Sect. 3 particularizes these definitions for
standard MVAR processes, derives the corresponding frequency domain measures of
coupling and causality, and discusses their interpretation; Sect. 4 proposes an extended
MVAR representation to be used in the presence of significant instantaneous correlations in
the observed process, whereby novel frequency domain causality measures are defined and
compared to the existing ones; Sect. 5 briefly discuss the practical application of the
measures on physiological time series; and Sect. 6 concludes the chapter.
2. Causality and coupling in multivariate processes
Let us consider M stationary stochastic processes y
m
, m=1, ,M. Without loss of generality
we assume that the processes are real-valued, defined at discrete time (y
m
={y

m
(n)}; e.g., are
sampled versions of the continuous time processes y
m
(t), taken at the times t
n
=nT, with T the
sampling period) and have zero mean (E[y
m
(n)]=0, where E[·] is the statistical expectation
operator). A MV closed loop process is defined as:
y
m
(n)=f
m
(Y
m
,Ý
l
|l≠m)+w
m
(n) , l,m=1, ,M, (1)
where f
m
is the function linking the set of the p past values of the m-th process, collected in
Y
m
={y
m
(n-1), ,y

m
(n-p)}, as well as the sets of the present and the p past values of all other
processes, collected in Ý
l
={y
l
(n),Y
l
}={y
l
(n),y
l
(n-1), ,y
l
(n-p)}, l≠m, to the present value y
m
(n),
and w
m
is a white noise process describing the error in the representation. Given two
processes y
i
and y
j
, i,j=1, ,M, different definitions of causality and coupling between the
processes may be defined as discussed in the following, and summarized in Table 1.
Biomedical Engineering Trends in Electronics, Communications and Software

406

Strictly causal MVAR
representation
Extended MVAR
representation

DIRECT

a) direct causality
y
j
→y
i

PDC, π
ij
( f )
PDC,
(
)
ij
f
π


b) extended direct causality
y
j →

y
i

-
ePDC,
χ
ij
( f )
c) direct coupling
y
i
↔y
j

PCoh, Π
ij
( f )
DIRECT+INDIRECT

a) causality y
j
⇒
y
i

DC, γ
ij
( f )
DC,
(
)
ij

f
γ


b) extended causality
y
j ⇒

y
i

-
eDC,
ξ
ij
( f )
c) coupling y
i
⇔
y
j

Coh, Γ
ij
( f )
Table 1. Frequency domain measures of causality and coupling between two processes y
i

and y
j

of a multivariate closed loop process. Note that causality and direct causality measure
lagged effects only, while extended causality and extended direct causality measure
combined instantaneous and lagged effects.
Denoting as Z
j
={Y
l
|l=1, ,M,l≠j} the set of the past values of all processes except y
j
, direct
causality from y
j
to y
i
, y
j
→y
i
, exists if the prediction of y
i
(n) based on Z
j
and Y
j
is better than
the prediction of y
i
(n) solely based on Z
j
. Causality from y

j
to y
i
, y
j
⇒ y
i
, exists if a cascade of
direct causality relations y
j
→y
m
···→y
i
occurs for at least one value m in the set (1,…,M); if
m=i or m=j causality reduces to direct causality. This last case is obvious for a bivariate
closed loop process (M=2), where only one definition exists and agrees with the notion of
Granger causality (Granger, 1969) involving only the relations between two processes. For
multivariate processes (M≥3) the definition of direct causality agrees with the notion of
prima facie cause introduced in (Granger, 1980); the definition of causality is a generalization
including also causal indirect effects between two processes, i.e., effects mediated by one or
more other processes in the MV closed loop.
While the definitions provided above are based on the exclusive consideration of lagged
effects from one series to another, the interactions modeled in (1) consider also the possible
instantaneous effects, i.e. effects which occur within the same lag. If we consider the directed
interaction from y
j
to y
i
, lagged causality (with lag k≥1) occurs if y

j
(n-k) is useful to predict
y
i
(n), while instantaneous causality (with lag k=0) occurs if y
j
(n) is useful to predict y
i
(n).
These two concepts may be combined together to provide extended causality definitions as
follows. Denoting as Z
ij
={Y
i
,Ý
l
|l=1, ,M,l≠j,l≠i} the set of the past values of y
i
and the present
and past values of all other processes except y
j
, extended direct causality from y
j
to y
i
, y
j
→

y

i
,
exists if the prediction of y
i
(n) based on Z
ij
and Ý
j
is better than the prediction of y
i
(n) solely
based on Z
ij
. Extended causality from y
j
to y
i
, y
j
⇒

y
i
, exists if a cascade of extended direct
causality relations y
j
→

y
m

··· →

y
i
occurs for at least one value m in the set (1,…,M); again, if
m=i or m=j extended causality reduces to extended direct causality.
Definitions of coupling between two processes are derived from the causality definitions as
follows. Direct coupling between y
i
and y
j
, y
i
↔y
j
, exists if y
i
→

y
m
and y
j
→

y
m
; while the most
obvious case is when m=i or m=j, two processes are considered as directly coupled also
when they both directly cause a third common process (m≠i, m≠j). Coupling between y

i
and
Multivariate Frequency Domain Analysis of Causal Interactions in Physiological Time Series

407
y
j
, y
i
⇔ y
j
, exists if y
m
⇒

y
i
and y
m
⇒

y
j
; again, coupling may arise when the of the two
processes causes the other (m=i or m=j), or when both processes are caused by other
common processes (m≠i, m≠j). Thus, the coupling definitions generalize the concept of
causality accounting for both forward and backward interactions between two processes.
An illustrative example of the described causality and coupling relations is reported in Fig.
1. In the diagrams, the set of interactions is represented with a network where nodes
correspond to processes and connecting arrows depict direct causality relations.

Fig. 1. Examples of networks of interacting processes exhibiting only lagged interactions (a)
and combined instantaneous and lagged interactions (b). Lagged and instantaneous effects
are depicted with solid and dashed arrows, respectively.
Fig. 1a shows a network of M=4 interacting processes in which only lagged effects from one
process to another are present. In this situation, extended causality reduces to causality due
to the absence of instantaneous effects. The direct causality relations imposed in the net are
y
1
→y
2
, y
2
→y
3
, y
3
→y
2
, and y
1
→y
4
. Since direct causality is a condition sufficient for causality,
we observe also y
1
⇒ y
2
, y

2
⇒ y
3
, y
3
⇒ y
2
, and y
1
⇒ y
4
; moreover, the cascade y
1
→y
2
→y
3

determines an indirect effect such that causality y
1
⇒ y
3
exists. Direct coupling follows from
direct causality, so that y
1
↔y
2
, y
2
↔y

3
, and y
1
↔y
4
, but is also caused by the common driving
exerted by y
1
and y
3
on y
2
, so that y
1
↔y
3
. Finally, coupling is present between each pair of
processes: y
1
⇔ y
2
, y
2
⇔
y
3
, y
1
⇔
y

4
, and y
1
⇔
y
3
result from the causality relations, while
y
2
⇔ y
4
and y
3
⇔ y
4
result from the common driving exerted by y
1
respectively on y
2
and y
4
,
and on y
3
and y
4
. In Fig. 1b, instantaneous effects are considered together with lagged ones.
In this case, direct causality occurs only when lagged effects are present, i.e., over the
directions y
1

→y
2
, y
3
→y
1
. Extended direct causality follows from lagged and/or
instantaneous direct causality, so that we have y
1
→

y
2
, y
2
→

y
3
, y
2
→

y
4
, and y
3
→

y

1
. As no
indirect lagged causality is present, causality follows exclusively from direct causality, i.e.
y
1
⇒ y
2
, y
3
⇒ y
1
. On the contrary, extended causality is observed very often because of the
existence of several cascades of instantaneous and/or lagged effects: we observe indeed
y
1
⇒

y
2
, y
1
⇒

y
3
, y
1
⇒

y

4
, y
2
⇒

y
1
, y
2
⇒

y
3
, y
2
⇒

y
4
, y
3
⇒

y
1
, y
3
⇒

y

2
, y
3
⇒

y
4
. Direct coupling
follows from extended direct causality: y
1
↔y
2
, y
2
↔y
3
, y
2
↔y
4
, y
3
↔y
1
(no common driving of
two processes on a third one is observed). Finally, coupling is detected between all pairs of
processes as the network is fully connected (i.e., there are no isolated groups of processes).
While causality definitions cannot be explored by means of conventional statistical
operators, the concepts of coupling and direct coupling may be quantified through standard
analysis of the correlation structure of the observed processes. Specifically, defining as

Y(n)=[y
1
(n)···y
M
(n)]
T
the observed M×1 vector process, as R(k)=E[Y(n)Y
T
(n-k)] its M×M
correlation matrix evaluated at lag k, and as P(k)=R(k)
-1
the inverse correlation matrix,
coupling y
i
⇔ y
j
and direct coupling y
i
↔y
j
are quantified at the time lag k respectively by the
correlation coefficient and the partial correlation coefficient (Whittaker, 1990):
Biomedical Engineering Trends in Electronics, Communications and Software

408

()
(
)
() ()

krkr
k
r
k
jjii
ij
ij
=
ρ
,
()
(
)
() ()
kpkp
k
p
k
jjii
ij
ij
−=
η
, (2)
where
r
ij
(k) and p
ij
(k) are the i-j elements of R(k) and P(k). The correlation and partial

correlation coefficients are normalized measures of the linear interdependence existing
between
y
i
(n) and y
j
(n-k), and of the linear interdependence between y
i
(n) and y
j
(n-k) after
removing the effects of all remaining processes. As such,
ρ
ij
and
η
ij
quantify the correlation
and the “direct correlation” (i.e., the correlation that cannot be accounted for by the
influence of any other process) between
y
i
and y
j
. To identify the frequency-domain
analogous of these two coefficients, we consider the spectral representation of the vector
process Y(
n), which is provided by the M×M spectral density matrix S( f ), defined as the
Fourier transform (FT) of the correlation matrix R(
k). The spectral matrix contains the

spectrum of
y
i
(n), S
ii
( f ), and the cross-spectrum between y
i
(n) and y
j
(n), S
ij
( f ), as diagonal
and off-diagonal terms, respectively (i,j=1,…,M). In analogy with the time domain
definitions, the spectral matrix and its inverse, P( f )=S( f )
-1
, are exploited to provide
frequency-domain measures of coupling and direct coupling, respectively through the
coherence (Coh) and the partial coherence (PCoh) functions (Bendat & Piersol, 1986):

()
(
)
() ()
fSfS
fS
f
jjii
ij
ij
=Γ ,

()
(
)
() ()
fPfP
fP
f
jjii
ij
ij
=Π , (3)
As the functions in (3) are complex-valued, their squared modulus is commonly used to
measure the strength of coupling and direct coupling in the frequency domain. Specifically,
the magnitude-squared Coh |
Γ
ij
( f )|
2
measures the strength of the linear, non-directed
interactions between the processes y
i
and y
j
as a function of frequency, being 0 in case of
uncoupling and 1 in case of full coupling. The squared PCoh |
Π
ij
( f )|
2
measures the

strength of the direct, non-directed interaction between y
i
and y
j
, i.e. the strength of the
interaction remaining after subtracting the effect of the remaining processes. We stress that,
due to the symmetrical nature of these measures, they cannot provide information about
causality; such an information may be extracted, as explained in the following, from the
coefficients of a parametric representation of the time series.
3. Causality and coupling in MVAR processes
3.1 Time domain definitions
The joint multivariate process Y(
n) can be represented as the output of a MV linear shift-
invariant filter (Kay, 1988):

() () ( )
∑
∞
−
∞
=
−=
k
knkn UHY
, (4)
where U(n)=[u
1
(n)···u
M
(n)]

T
is a vector of M zero-mean input processes and H(k) is the M×M
filter impulse response matrix. A particular case of the general model in (4), extensively
used in time series analysis, is the MV autoregressive (MVAR) model (Kay, 1988):

() ()( ) ()
nknkn
p
k
UYAY +−=
∑
=
1
, (5)
Multivariate Frequency Domain Analysis of Causal Interactions in Physiological Time Series

409
where
p is the model order, defining the maximum lag used to quantify interactions. The
input process
U(n), also called innovation process, is assumed to be composed of white and
uncorrelated noises; this means that the correlation matrix of
U(n), R
U
(k)=E[U(n)U
T
(n-k)], is
zero for each lag
k>0, while it is equal to the covariance matrix Σ=cov(U(n)) for k=0. One
major benefit of the representation in (5) is that it allows to interpret properties of the joint

description of the processes
y
m
(n) —like coupling or causality— in terms of the estimated
coefficients
A(k). In fact, the i-j element of A(k), a
ij
(k), quantifies the causal linear interaction
effect occurring at lag
k from y
j
to y
i
. As a consequence, the definitions of causality and
coupling provided above for a general closed-loop MV process can be specified for a MVAR
process in terms of the off-diagonal elements of
A(k) as follows: y
j
→y
i
if a
ij
(k)≠0 for at least
one
k=1,…,p; y
j
⇒ y
i
if
(

)
0
1
≠
−
smm
ka
ss
for at least a set of L≥2 values for m
s
(with m
0
=j, m
L-1
=i)
and a set of lags
k
0
, ,k
L-1
with values in (1, ,p); y
i
↔y
j
if a
mi
(k
1
)≠0 for at least one k
1

or a
mj
(k
2
)≠0
for at least one
k
2
; y
i
⇔
y
j
if
(
)
0
1
≠
−
smm
ka
ss
for at least a set of L≥2 values for m
s
(either with
m
0
=m, m
L-1

=i or with m
0
=m, m
L-1
=j) and a set of lags k
0
, ,k
L-1
. Thus, causality and coupling
relations are found when the pathway relevant to the interaction is active, i.e., is described
by nonzero coefficients in
A. Note that the extended definitions of causality and direct
causality cannot be tested from the coefficients of the MVAR model (5), as the model does
not describe instantaneous interactions. We refer to Sect. 3.3 to see how the MVAR
coefficients may be related to causality and coupling effects in an illustrative example.
3.2 Frequency domain definitions
The spectral representation of a MVAR process is derived considering the FT of the
representations in (4) and (5), which yield respectively the equations
Y( f )=H( f )U( f ) and
Y( f )=A( f )Y( f )+U( f ), where Y( f ) and U( f ) are the FTs of Y(n) and U(n), and the M×M
transfer matrix and coefficient matrix are defined in the frequency domain as:

(
)
()
kTfj
k
ekf
π
2−

∞
−∞=
Σ
= HH ,
(
)
()
kTfj
p
k
ekf
π
2
1
−
=
Σ
= AA . (6)
Comparing the two spectral representations, it is easy to show that the coefficient and
transfer matrices are linked by:
H( f )=[I-A( f )]
-1
=Ā( f )
-1
. This relation is useful to draw the
connection between the cross-spectral density matrix
S( f ) and its inverse P( f ), as well as to
derive frequency domain estimates of coupling and causality in terms of the MVAR
representation. Indeed, the following factorizations hold for a MVAR process (Kay, 1988):
S( f )=H( f )ΣH

H
( f ) , P( f )=Ā
H
( f )Σ
-1
Ā( f ), (7)
where the superscript
H
stands for the Hermitian transpose. The (i-j)th elements of S( f ) and
P( f ) can be represented in the compact form:

(
)
(
)
(
)
fffS
jiij
H
Σhh= ,
(
)
(
)
(
)
fffP
jiij
aΣa

1H −
= , (8)
where
h
i
( f ) is the i-th row of the transfer matrix (H( f )=[h
1
( f )···h
M
( f )]
T
) and ā
i
( f ) is the i-th
column of the coefficient matrix (Ā( f )=[ā
1
( f )···ā
M
( f )]). Under the assumption that the
input white noises are uncorrelated even at lag zero, their covariance cov(
U(n)) reduces to
the diagonal matrix
Σ=diag(
σ
2
i
), and its inverse to the matrix Σ
-1
=diag(1/
σ

2
i
) which is
diagonal as well (
σ
2
i
is the variance of u
i
). In this specific case, (8) factorize into:
Biomedical Engineering Trends in Electronics, Communications and Software

410

() () ()
∑
=
=
M
m
*
jmimmij
fHfHfS
1
2
σ
,
() () ()
∑
=

=
M
m
mj
*
mi
m
ij
fffP
1
2
1
AA
σ
(9)
The usefulness of the factorizations in (9) is in the fact that they allow to decompose the
frequency domain measures of coupling and direct coupling previously defined into terms
eliciting the directional information from one process to another. Substituting (8) and (9)
into (3), the Coh between
y
i
and y
j
can be factored as:

()
(
)
(
)

() ()
() ()
(
)
()
(
)
()
() ()
∑∑
==
===Γ
M
m
*
jmim
M
m
jj
*
jmm
ii
imm
jjii
ji
ij
ff
fS
fH
fS

fH
ffff
ff
f
11
HH
H
γγ
σ
σ
ΣhhΣhh
Σhh
, (10)
where the last term contains the so-called directed coherence (DC). Thus, the DC from y
j
to y
i
is
defined as (Baccala et al., 1998):

()
(
)
()
∑
=
=
M
m
imm

ijj
ij
fH
fH
f
1
2
2
σ
σ
γ
. (11)
Note that the directed transfer function (DTF) defined in (Kaminski & Blinowska, 1991) is a
particularization of the DC in which all input variances are all equal (σ
2
1
=σ
2
2
=···=σ
2
M
) so
that they cancel each other in (11). The factorization in (10) justifies the term DC, as
γ
ij
( f ) can
be interpreted as a measure of the influence of
y
j

onto y
i
, as opposed to
γ
ji
( f ) which
measures the interaction occurring over the opposite direction from
y
i
to y
j
. Further
interpretation of the DC in terms of coupling strength is achieved considering its
normalization properties:

() ()
110
1
2
2
=≤≤
∑
=
M
m
imij
f,f
γγ
. (12)
The inequality in (12) indicates that the squared DC |

γ
ij
( f )|
2
measures a normalized
coupling strength, being 0 in the absence of directed coupling from y
j
to y
i
at the frequency f,
and 1 in the presence of full coupling. The equality indicates that |
γ
ij
( f )|
2
measures the
coupling strength from
y
j
to y
i
as the normalized proportion of S
ii
( f ) which is due to y
j
, i.e. is
transferred from u
j
via the transfer function H
ij

( f ). Indeed, combining (9) and (12) it is easy
to show that the spectrum of the process y
i
may be decomposed as:

() () () () ()
fSffS,fSfS
iiimm|i
M
m
m|iii
2
1
γ
==
∑
=
. (13)
where S
i|m
( f ) is the part of S
ii
( f ) due to y
m
; S
i|i
( f ) measures the part of S
ii
( f ) due to none of
the other processes, which is quantified in normalized units by the squared DC |

γ
ii
( f )|
2
.
Note that the useful decomposition in (13) does not hold for the DTF, unless all input
variances are equal to each other so that the DC reduces to the DTF. For this reason, in the
following we will consider the DC only, as it provides a similar, but more general and
interpretable in terms of power content, measure of frequency domain causality.
In a similar way to that followed to decompose the Coh, the PCoh defined in (3) can be
factored, using (8) and (9), as:
Multivariate Frequency Domain Analysis of Causal Interactions in Physiological Time Series

411
()
() ()
() () () ()
()
()
()
()
() ()
∑∑
==
−−
−
===Π
M
m
*

mimj
M
mii
*
mi
m
jj
mj
m
jjii
ji
ij
ff
fP
fA
fP
fA
ffff
ff
f
11
1H1H
1H
11
ππ
σσ
aΣaaΣa
aΣa
, (14)
where the last term contains the

partial directed coherence (PDC) functions. The PDC from y
j
to
y
i
is thus defined as (Baccala et al., 2007):

()
()
()
∑
=
=
M
m
mj
m
ij
i
ij
fA
fA
f
1
2
2
1
1
σ
σ

π
. (15)
As suggested by the factorization in (14), the PDC extracts the directional information from
the PCoh, and is thus a measure of the direct directed interaction occurring from y
j
to y
i
at
the frequency f. The normalization properties for the squared modulus of the PDC are:

() ()
110
1
22
=≤≤
∑
=
M
m
mjij
f,f
ππ
, (16)
suggesting that |
π
ij
( f )|
2
quantifies the interaction from y
j

to y
i
as the normalized proportion
of P
jj
( f ) which is which is sent to y
i
, via the coefficients Ā
ji
( f ). Indeed, we have that:

() () () () ()
fPffP,fPfP
jjmjmj
M
m
mjjj
2
1
π
==
→
=
→
∑
, (17)
where
P
j→m
( f ) is the part of P

jj
( f ) sent to y
m
; in particular, P
j→j
( f ) measures the part of P
jj
( f )
which is not sent to the other processes, and is expressed in normalized terms by the
squared PDC |
π
jj
( f )|
2
. The quantity which we denote as PDC was named “generalized
PDC” in (Baccala et al., 2007), while the original version of the PDC (Baccala & Sameshima,
2001) was not including inner normalization by the input noise variances. Our definition
(15) follows directly from the decomposition in (14); besides, this definition shares with the
Coh, PCoh and DC functions the desirable property of scale-invariance, contrary to the
original PDC that may be affected by different amplitudes for the considered signals.
Although both DC and PDC may be regarded as frequency domain descriptors of causality,
there are important differences between these two estimators. First, as the DC and the PDC
are factors in the decomposition of Coh and PCoh, respectively, they measure causality and
direct causality in the frequency domain. In fact, the PDC
π
ij
( f ) is nonzero if and only if
direct causality
y
j

→y
i
exists, because the numerator of (15) contains, with i≠j, the term
()
fA
ij
, which is nonzero only when a
ij
(k)≠0 for some k and is uniformly zero when a
ij
(k)=0
for each
k. As to the DC, one can show that, expanding H( f )=Ā( f )
-1
as a geometric series,
the transfer function
H
ij
( f ) contains a sum of terms each one related to one of the (direct or
indirect) transfer paths connecting
y
j
to y
i
(Eichler, 2006). Hence, the DC
γ
ij
( f ) is nonzero
whenever any path connecting
y

j
to y
i
is significant, i.e., when causality y
j
⇒
y
i
occurs.
Another important difference between DC and PDC is in the normalization: as seen in (12)
and in (16),
γ
ij
( f ) is normalized with respect to the structure that receives the signal, while
π
ij
( f ) is normalized with respect to the structure that sends the signal. Summarizing, we can
Biomedical Engineering Trends in Electronics, Communications and Software

412
state that the DC measures causality as the amount of information flowing from y
j
to y
i

through all (direct and indirect) transfer pathways, relative to the total inflow entering the
structure at which
y
i
is measured; the PDC measures direct causality as the amount of

information flowing from
y
j
to y
i
through the direct transfer pathway only, relative to the
total outflow leaving the structure at which
y
j
is measured. We note that this dual
interpretation highlights advantages and disadvantages of both measures. The DC has a
meaningful physical interpretation as it measures causality as the amount of signal power
transferred from one process to another, but cannot distinguish between direct and indirect
causal effects measured in the frequency domain. Conversely, the PDC reflects clearly the
underlying interaction structure as it provides a one-to-one representation of direct
causality, but is hardly useful as a quantitative measure because its magnitude quantifies
the information flow through the inverse spectral matrix elements (which do not find easy
interpretation in terms of power spectral density).
3.3 Theoretical example
To discuss the properties and compare the behavior of the frequency domain measures of
causality and coupling summarized in Table 1, we consider the MVAR vector process of
order
p=2, composed by M=4 processes, generated by the equations:

()
(
)
(
)
(

)
(
)
()() ()()
()
()
() () () ()()
() ( ) ()
⎪
⎪
⎩
⎪
⎪
⎨
⎧
+−=
+−+−+−−−=
+−+−=
+−−−=
nunyny
nuny.ny.nynyfcosny
nuny.nyny
nunynyfcosny
414
3223
2
333
2312
11
2

111
2
2501502122
1501
2122
ρπρ
ρπρ
, (18)
with
ρ
=0.9, f
1
=0.1 and f
3
=0.3, where the inputs u
i
(n), i=1,2,3, are fully uncorrelated and with
variance
σ
2
i
. The process (18) is one of the possible MVAR realizations of the diagram of Fig.
1a. The coupling and causality relations emerging from the diagram, discussed in Sect. 2,
can be interpreted here in terms of the MVAR coefficients set in (18). In fact, the nonzero off-
diagonal values of the coefficient matrix (
a
21
(1)=1, a
23
(1)=0.5, a

32
(1)=0.5, a
32
(2)=0.5, a
41
(2)=1)
determine direct causality and causality among the processes —and consequently direct
coupling and coupling— in agreement with the definitions particularized at the end of Sect.
3.1. For instance,
a
21
(1) and a
32
(2)=0.5 determine direct causality y
1
→y
2
and y
2
→y
3
as well as
causality
y
1
⇒ y
2
, y
2
⇒ y

3
(direct interaction) and y
1
⇒ y
3
(indirect interaction). The diagonal
values of the coefficient matrix determine autonomous oscillations in the processes. Indeed,
the values set for
a
ii
(k), a
ii
(1)=2
ρ
cos(2
π
f
i
), a
ii
(2)=-
ρ
2
, generate complex-conjugate poles with
modulus
ρ
and phases ±2
π
f
i

for the process y
i
(the sampling period is implicitly assumed to
be
T=1). In this case, narrow-band oscillations at 0.1 Hz and 0.2 Hz are set for y
1
and y
3
.
The trends of spectral and cross-spectral density functions are reported in Fig. 2.
The spectra of the four processes, reported as diagonal plots in Fig. 2a (black), exhibit clear
peaks at the frequency of the two imposed oscillations: the peaks at ~0.1 Hz and ~0.3 Hz are
dominant for
y
1
and y
3
, respectively, and appear also in the spectra of the remaining
processes according to the imposed causal information transfer. The inverse spectra,
computed as the diagonal elements of the inverse spectral matrix
P( f ), are also reported as
diagonal plots in Fig. 2b (black). Off diagonal plots of Fig. 2a and Fig. 2b depict respectively
the trends of the squared magnitudes of Coh and PCoh; note the symmetry of the two
Multivariate Frequency Domain Analysis of Causal Interactions in Physiological Time Series

413
functions (Γ
ij
( f )=Γ*
ji

( f ), Π
ij
( f )=Π*
ji
( f )), reflecting the fact that they measure coupling and
direct coupling but cannot account for directionality of the considered interaction. As
expected, Coh is nonzero for each pair of processes, thus measuring the full connectivity of
the considered network. PCoh is clearly nonzero whenever a direct coupling relation exists
(
y
1
↔y
2
, y
2
↔y
3
, y
1
↔y
4
, y
1
↔y
3
), and is uniformly zero between y
2
and y
4
and between y

3
and
y
4
where no direct coupling is present.
Figs. 3 and 4 depict the decomposition of spectra and inverse spectra, as well as the trends of
DC and PDC functions resulting from these decompositions.
Fig. 3a provides a graphical representation of (13), showing how the spectrum of each
process can be decomposed into power contributions related to all processes; normalizing
these contributions one gets the squared modulus of the DC, as depicted in Fig. 3b. In the
example, the spectrum of
y
1
is decomposed in one part only, deriving from the same
process. This indicates that none part of the power of
y
1
is due to the other processes. The
absence of external contributions is reflected by the null profiles of the DC from
y
1
to y
2
, to y
3

and to
y
4
seen in Fig. 3b; as a result, the squared DC |

γ
11
( f )|
2
has a flat unitary profile. On
the contrary, the decompositions of
y
i
, with i=2,3,4, results in contributions from the other
processes, so that the squared DC |
γ
ij
( f ) |
2
is nonzero for some j≠i, and the squared DC
|
γ
ii
( f ) |
2
is not uniformly equal to 1 as a result of the normalization condition. In particular,
we observe that the power of the peak at
f
1
=0.1 Hz is entirely due to y
1
for all processes,
determining very high values of the squared DC in the first column of the matrix plot in Fig.
3b, i.e., |
γ

i1
( f
1
)|
2
=1; this behavior represents in the frequency domain the causality relations
imposed from
y
1
to all other processes. The remaining two causality relations, relevant to the
bidirectional interaction between
y
2
and y
3
, concern the oscillation at f
2
=0.3 Hz, which is




S
ii
( f ): spectrum of the process y
i
; P
ii
( f ): inverse spectrum of y
i

; Γ
ij
( f ): coherence between y
j
and
y
i
; Π
ij
( f ): partial coherence between y
j
and y
i
.
Fig. 2. Spectral functions and frequency domain coupling measures for the theoretical
example (18).