Authors: Tomáš Havel;  Jiří Balach
Authors‘ workplace: Faculty of Electrical Engineering, CTU in Prague, Czech Republic
Published in: Lékař a technika - Clinician and Technology No. 2, 2012, 42, 92-95
Category: Conference YBERC 2012


In this paper we present our implementation of correlation analysis method used for detecting epileptic seizures during their onset zone. We used matrices of correlation coefficients and their time course to classify seizure. We also developed effective method to reduce number of features. Implemented algorithms were tested on our own dataset acquired in cooperation with the neurological department of FN Motol. The results show this approach as reasonable but the accuracy must be significantly improved before use in practice.

scalp EEG, onset seizure detection, epilepsy, correlation analysis, ictal SPECT


Surgical treatment is mostly indicated in patients with pharmacoresistant epilepsy. In this case, additional examinations are indicated especially various tomographic scans. One of them is also subtraction of SPECT. This exam displays differences in cerebral perfusion during epileptic seizure. Subtraction image is obtained from ictal and interictal SPECT. These scans use a contrast agent (the radiopharmaceutical, the RP) which are totally recaptured in brain tissue in approximately 30 seconds [1].

When an epileptic patient needs an ictal SPECT scan examination the medical nurse or doctor has to be with the patient all the time, mostly for hours and focus on his behavior for early recognition of the seizure. When the staff realizes the seizure the RP has to be injected. The seizure has to be detected early because if the RP is delivered late, the scan will be less accurate or it can even image a false location instead of a prime epileptic focus.

Our interest is to develop a robust system which will help to the medical personnel to decide whether the patient has the seizure. This system should recognize the seizure from changes in scalp EEG signal and warn the doctor [7].

If the system works robustly enough, the doctor would not have to continuously monitor the EEG and come only when it is necessary. The most advanced idea is to design automatic system which integrates EEG monitor, analyzer machine and linear infusion pump capable to work autonomously or just with doctor confirmation of the RP injection.

Possibilities od seizure detection

Detection of epileptic seizures from EEG is the determination of time zone in which occurs instead interictal physiological EEG course signal, which can be described as ictal pathophysiological course.

There are many different approaches how to determine the seizure from EEG signal. Today's popular methods of seizure detection usually use some type of signal transformation like Fourier, Wavelet etc. These methods are called time-frequency methods after frequency images calculated for time segments [8].

Good results were also achieved by time feature extraction and amplitude-frequency approaches oriented to single or multichannel [2] signal processing [3, 9]. Correlation method can be classify as multichannel time feature extraction technique and were already more-less successfully used for the EEG signal processing. Their usage is mostly oriented for epileptogenic source localization [4] in intracranial EEG inspection.

Another big class of seizure detection methods is pattern recognition. These methods use various analysis and react when some pattern typical for seizure occurs [6].

Description of correlation method

When we use correlation analysis for seizure onset time localization at first we have to segment signals to separate parts and assign timestamps to these parts. We use a windowing method similar to sliding window. It is important to choose right length of segments, because too short segment contains poor information and long segment does not allow good time resolution. To improve time precision we can use overlap of windows but even this method does not solve all problems. When the overlap is too large, computational requirements are unreasonably high and furthermore results are blurred. Segments of multichannel EEG are sets of n short-time signals and one technique to get their mutual information is to calculate cross correlation.

The cross correlation of n EEG channels is calculated in each time window. It gives n-by-n matrix of correlation coefficients (Fig. 1). These coefficients should describe dependence (similarity) and in this application especially synchronicity of signals with zero lag. The final product is map of channel similarities in discrete time steps. Those characteristics change in time. Assuming the neural activity of the brain cells is during seizure abnormally synchronized, this pattern should appear in the correlation maps. The matrix of correlation coefficients is diagonally symmetrical and coefficients on diagonal correspond to autocorrelation which are equal to one. It follows n (n −1 )/ 2 coefficients useful for further processing. It is still huge amount of information from which we are trying to extract some results. 

Fig. 1: Correlation map (matrix of correlation coefficients) from ictal zone.
Fig. 1: Correlation map (matrix of correlation coefficients) from ictal zone.

We are using further algorithmic processing to evaluate the time course of correlation maps. For reduction of multivariate data we can calculate eigenvalues of each correlation matrix [5]. The advantage is that the distribution of the eigenvalues λ and their corresponding eigenvectors v of correlation matrix C is directly related to the correlation structure of the multichannel EEG. The eigenvalues are computed by solving the equation (1). 

The eigenvectors form a new orthogonal basis of the n-dimensional vector space spanned by the n EEG channels. Orthogonality ensures uncorrelated state. In the n-dimensional vector space spanned by the original EEG channels, the eigenvector vmax is associated with the largest eigenvalue λmax. Due n orthogonal eigenvectors, C may be diagonalized (linearly transformed into a matrix that contains elements only on its main diagonal). The diagonal elements become the eigenvalues λ and their amplitude is proportional to the amount of correlation in the direction of their associated eigenvectors. Under linear transformations the sum of the diagonal elements of a matrix remains unchanged. The sum of the eigenvalues λ then must be equal to the sum of the diagonal elements of the original correlation matrix C and it is equal to number of EEG channels n.

The equation (2) must always be respected and consequently, if some eigenvalues increase another one has to decrease to keep the sum constant. Ascending order of eigenvalues called spectrum of the correlation matrix. It can be visualized for each time window in one map (Fig. 2). Illustration of more and less correlated segments are shown on artificialy modified EEG signal in figure (Fig. 3).

Fig. 2: Spectrum map of the correlation matrix. Seizure detected in onset zone .
Fig. 2: Spectrum map of the correlation matrix. Seizure detected in onset zone .

Fig. 3: Spectrum map of the correlation matrix. This map demonstrate effect of more correlated segment from window 37 to 48 and less correlated segment from window 85 to 126. Signals were artificially modified.
Fig. 3: Spectrum map of the correlation matrix. This map demonstrate effect of more correlated segment from window 37 to 48 and less correlated segment from window 85 to 126. Signals were artificially modified.

Evaluation method uses eigenvalues (spectrum of the correlation matrix) is very useful but each time moment is still described by n features vector. We wanted to found some feature which can represent this vector and approximately specify the situation in time with just one parameter. We tried few possibilities and found one as usable and effective. This method count parameter called s:

This parameter s represents distance of the state from the coordinate origin and can reach values between √n and n. Static value of this parameter doesn't bring a lot of information but dynamic changes can demonstrate evolution of correlation between channels in time. When the value is equal to √n, the channels are absolutely orthogonal and when is equal to n, they are absolutely correlated. With this knowledge we can use this waveform from interictal segment of signal to set a base line of parameter s of EEG signal. Then we can also set some threshold value for seizure detection based on statistical distribution of interictal signal (Fig. 5).

Fig. 4: Waveform of filtered parameter s. Seizure detected in onset zone
Fig. 4: Waveform of filtered parameter s. Seizure detected in onset zone


We implemented algorithms described in previous chapter with the following parameters. Segmentation of signal was realized by windowing technique to segments length 3 s with 3/4 overlap. This setting provides time resolution 0.75 s which we consider as sufficient. We filtered the signals to few frequency bands and calculate cross correlation for each band separately. We used following bands 0-2 Hz, 2-10Hz, 10-20Hz, 20-60Hz and 60-100Hz. Advantege of this technique is higher sensitivity of some frequency band because every patient has different symptoms in ictal EEG. This method proved as good approach because without filtration in most cases correlation maps doesn't show big changes. Correlation maps were evaluated by method uses calculation of eigenvalues followed by calculation of parameter s (3). We experimentally discovered that better results are achieved by additional filtration of this waveform with moving average filter with length 3 (Fig. 4). This course is classified by dynamical threshold.

Experimental verification and results

For experimental verification of implemented algorithm were used scalp EEG records with 18 bipolar channels in bipolar longitudinal montage (“Double banana”). All records have sample frequency 200 Hz.

We early determine the short records as useless for this technique. It is needed at least 5 minutes prior to seizure onset. That unfortunately limits our resources. We had 11 records complying with these requirements obtained from 6 different patients. Patients were in age from childhood to adult and suffered with various epileptic disease. During implementation we found the time course of correlation maps of unfiltered signals does not correspond to the seizure zones annotated by doctors. When we use signals filtered to bands described above we found the results from some bands as useful for seizure onset detection. Positive results were achieved especially with filtered signals in band 60-100Hz. The interictal zone used for estimation of base line and it's threshold were 30 – 130 s of EEG record. Onset zone was chosen as 20 second after doctor annotation.

The obtained results are as follows: 6 seizures were detected in onset zone, 2 seizures were detected during seizure after onset zone, in 1 record was false detection before seizure starts and in 2 records were seizures not detected. These results show the algorithm was successful in more than half the cases.

The dataset is not big enough for statistically significant conclusions. Further tests with larger dataset should be carried. Currently, we can recommend use of this type of analysis for developing systems for onset seizure detection in combination with another analysis method.


From the beginning was the method conceived for detection onset zone of epileptic seizure from scalp multichannel EEG record although based on the methods more often used in the analysis of intracranial EEG. Our motivation was make own implementation of this method and experimentally verify it's function on available dateset of EEG records from patients of different ages and epileptic disease. The results show this implementation of correlation analysis successful in 6 of 11 cases as onset seizure detector. Failures were caused by lack of change of synchronicity in EEG records or in case of false detection by spontaneous synchronization in interictal zone. We assume better results would achieve on a more limited group of patients and/or epileptic disease.

We verified our implementation of correlation analysis on our test dataset. Although the results are not very convincing, this method is applicable to the construction of onset seizure detection system.


This work has been supported by the grants IGA NT11460-4/2010 Intracranial EEG signal processing; epileptogenic zone identification in non-lesional refractory epilepsy patients, IGA NT13357-4/2012, SGS 10/272/OHK4/3T/13 Analysis of intracranial EEG recording, and resesearch program MSM6840770012 Transdisciplinary Research in Biomedical Engineering.

Ing. Tomáš Havel

Department of Circuit Theory

Faculty of Electrical Engineering

Czech Technical University in Prague

Technická 2, 166 27 Praha 6, Czech Republic


Phone: +420 604 277 615


[1] Funkčně zobrazovací vyšetření. Brázdil, M. a Marusič, P. Epilepsie temporálního laloku. Vyd. 1. Praha: Triton, 2006, s. 172-187. ISBN 80-7254-836-0.

[2] Duun-Henriksen, J., Kjaer, T. W., Madsen, R. E., Remvig, L. S., Thomsen, C. E., Sorensen, H. B. D. Channel selection for automatic seizure detection, Clinical Neurophysiology, Volume 123, Issue 1, January 2012, Pages 84-92, ISSN 1388-2457

[3] Janča, R., Čmejla, R., Jahodová, A. Rules for Spike Detection in Multichanel Intracranial Electroencephalography 19th Annual Conference Proceedings Technical Computing Prague 2011 Vydavatelství VŠCHT Praha, 2011, pp. -

[4] Schindler, K., Leung, H., Elger, C. E., & Lehnertz, K. (2007). Assessing seizure dynamics by analysing the correlation structure of multichannel intracranial EEG. Brain: A journal of neurology, 130(Pt 1), 65-77. Oxford Univ Press. Retrieved from

[5] Müller, M.; Baier, G.; Galka, A.; Stephani, U. & Muhle, H. Detection and characterization of changes of the correlation structure in multivariate time series Phys. Rev. E, American Physical Society, 2005, 71, 046116, DOI: 10.1103/PhysRevE.71.046116.

[6] Gotman, J. Automatic recognition of epileptic seizures in the EEG. Electroencephalography and Clinical Neurophysiology. 1982, roč. 54, č. 5, s. 530-540. ISSN 00134694. DOI: 10.1016/0013-4694(82)90038-4.

[7] Hao Qu, Gotman, J. A patient-specific algorithm for the detection of seizure onset in long-term EEG monitoring: possible use as a warning device. IEEE Transactions on Biomedical Engineering. 1997, roč. 44, č. 2, s. 115-122. ISSN 00189294. DOI: 10.1109/10.552241. Retrieved from Arnumber=552241

[8] Iscan, Z., Dokur, Z., Demiralp, T. Classification of electroencephalogram signals with combined time and frequency features. Expert Systems with Applications. 2011, roč. 38, č. 8, s. 10499-10505. ISSN 09574174. DOI: 10.1016/j.eswa.2011.02.110. Retrieved from

[9] Mcsharry, P. E., He, T., Smith, L. A., Tarassenko, L. Linear and non-linear methods for automatic seizure detection in scalp electro-encephalogram recordings. Medical. 2002, roč. 40, č. 4, s. 447-461. ISSN 0140-0118. DOI: 10.1007/BF02345078. Retrieved from


Article was published in

The Clinician and Technology Journal

Issue 2

2012 Issue 2

Most read in this issue
Forgotten password

Don‘t have an account?  Create new account

Forgotten password

Enter the email address that you registered with. We will send you instructions on how to set a new password.


Don‘t have an account?  Create new account