Epileptic Seizure Prediction by
Extracting Global and Local Features
Mohammad Zavid Parvez and Manoranjan Paul
School of Computing and Mathematics
Charles Sturt University, Australia
Motivation
•
It is difficult to provide a good balance between high seizure prediction accuracy
(100%) in advance with low false prediction rate (FPR) calculated in hour time unit
using all patients by a prediction algorithm.
• For a given seizure prediction horizon (SPH), it is also difficult to provide seizure
prediction performance above the chance level for all patients by a particular
method.
Therefore, more research should be conducted in this scope to achieve better
accuracy for the advanced prediction with low FPR.
Two features, mean energy concentration ratio
(MECR) and energy of cost function of weighted
fluctuation and deviation (ECFD), are considered
to predict seizures as follows:
and
Proposed Method
 In this paper, an effective epileptic seizure prediction method is proposed which
exploits undulated global and local features together with a regularization technique
to predict the seizure onset in advanced.
 In general, pre-processing, features extraction, classification, regularization (i.e.,
post-processing) and decision function are the possible steps for predicting seizure
in advance from Electroencephalogram(EEG) signals.
Fig. 4. Original EEG signals comprises interictal and preictal/ictal
(prIc) periods.
Fig. 5. Mean energy concentration ratio (MECR) of interictal and
preictal/ictal periods (prIc) using six channels.
Fig. 1. Generic block diagram of a seizure prediction process.
•
•
•
•
The proposed method is considered with a certain range of artifacts tolerance
without filtering technique.
Global feature is extracted using phase correlation between two consecutive epochs
of a signal and local feature is extracted using a weighted cost function comprising
fluctuation and deviation within an epoch.
The popular support vector machine (SVM) classifier is used to classify the interictal,
preictal, and ictal signals.
To purify the classified output, a two-step post-processing regularization technique is
also applied for the final output.
Fig. 6. Energy of cost function of fluctuation and deviation (ECFD)
of interictal and preictal/ictal (prIc) periods using six channels.
Fig. 7. Classified results of each epoch of interictal and
preictal/ictal (prIc) periods.
Fig. 8. Decision on classified values after windowing of interictal
and preictal/ictal (i.e., prIc) periods.
Fig. 2. Extracted global feature using phase correlation.
Experimental Results
The proposed method
outperforms six existing
relevant state-of-the-art
methods
considering
the balance between
the prediction accuracy
(%) and FPR.
Fig. 3. Extracted local feature using fluctuation and deviation.
References
1.
2.
3.
4.
5.
6.
J.R. Williamson, et al. (2012). Seizure prediction using EEG spatiotemporal correlation structure. Epilepsy Behav., 25 (2), 230–238.
L. Chisci, et al. (2010). Real-Time Epileptic Seizure Prediction Using AR Models and SVMs. IEEE Trans. BME, 57(5), 1124-1132.
P. Mirowski, et al. (2009). Classification of patterns of EEG synchronization for seizure prediction. Clin. Neurophysiol.
Y. Park, et al. (2011). Seizure prediction with spectral power of EEG using cost-sensitive SVM. Epilepsia, 52(10), 1761-1770.
S. Li, et al. (2013). Seizure Prediction Using Spike Rate of Intracranial EEG. IEEE Trans. Neural Systems and Rehabilitation, 21(6), 880-886.
M. Z. Parvez, and M. Paul, “Epileptic Seizure Prediction by Exploiting Spatiotemporal Relationship of EEG Signals using Phase Correlation,”
IEEE Trans. Neural Systems and Rehabilitation, 2015.
Contact details:
Mohammad Zavid Parvez
Phone: +61 2 6338 4322
Email: [email protected]
www.csu.edu.au