پیش بینی تشنج با استفاده از خصوصیات محلی و کلی نوسانی Seizure Prediction using Undulated Global and Local Features

EIZURE is a sudden surge of electrical activity of the brain affecting more than 65 million individuals (i.e. 1%) worldwide [1]. Approximately 325 million people experience a seizure within their life time [2]. During seizure, the brain cannot perform normal tasks; therefore, people may restriction and abnormal activity in movement, behavior, awareness, and sensation. Epilepsy is spontaneously recurrent seizures. Seizure causes many injuries such as submersion, burns, accidents, and more seriously, death. However, it is possible to prevent these unwanted situations by timely and correct prediction of epilepsy before the actual seizure onset. Electroencephalogram (EEG) is a wellaccepted tool for analyzing seizure [3]-[28]. EEG can measure electrical activity of the brain through multiple electrodes placed on the scalp Manuscript submitted for review July 10, 2015.This work was supported in part by the CM3 Machine Learning Research Centre, Charles Sturt University, Australia. The authors are with CM3, Charles Sturt University, Australia (e-mail: {mparvez; mpaul}@csu.edu.au). [8]. A significant amount of research of seizure prediction including [23]-[28] has been conducted. Williamson et al. [23] proposed a seizure prediction method based on spatiotemporal features. The experimental results provide 85% accuracy with false positive rate (FPR) of 0.03/h using 19 patients from a total of 21 patients using the benchmark data set [29]. Chisci et al. [24] also proposed a prediction method using an autoregressive model and support vector machine (SVM). The prediction accuracy (PA) was 100% with FPR of 0.41/h using only 9 patients from the same data set [29]. Mirowski et al. [25] proposed another method based on bivariate features such as cross-correlation, nonlinear interdependence, and dynamic entrainment using the data set in [29] where the results provided 71% accuracy with zero FPR using 15 of 21 patients. Park et al. [2] proposed a technique using linear features of spectralpower and non-linear classifier considering 18 of 21 patients that provided 94.4% accuracy with FPR of 0.20/h using the data set [29]. Li et al. [26] employed the spike rate using a morphological filter and obtained 75.8% PA with FPR of 0.09/h using all 21 patients from the data set [29]. Moghim et al. [27] proposed a seizure prediction technique in advance using different statistical features by preictal period relabeling of the EEG signals. They obtained high accuracy (i.e. 96.30%) for prediction between 1 and 6 minutes in advance using the data set [29]. Rasekhi et al. [28] proposed a seizure prediction technique based on linear univariate features by providing 73.9% PA with FPR of 0.15/h using another data set. It is difficult to achieve a good balance by using a prediction algorithm between high PA (100%) with low FPR, using all patients. Moreover, for a given seizure prediction horizon (SPH), it is also difficult to achieve prediction performance above the chance level for all patients by a particular method [11]. The non-abruptness phenomena and inconsistency of the signals along with different brain locations, patient-age, patient-sex, and seizure-type are the challenging issues that affect the consistency of performance in terms of advanced PA and false alarms by the existing methods using all types of patients. Therefore, more research should be conducted within this scope to achieve better accuracy for advanced prediction with low FPR. When an EEG signal is captured from a patient, it may have different types of periods such as interictal, preictal, and ictal periods, in that order. Thus, for the advanced prediction of an ictal state, the transition between interictal and preictal periods needs to determine, as EEG is a supreme advantage in studying transient neuronal activity [30]. How early a preictal period associated with an ictal period is determined from the interictal period by a technique indicates its advanced prediction performance accuracy. To determine the transition between interictal and preictal periods, an EEG signal needs to be processed by dividing the signal into a number of epochs (i.e. a specified time-window). Sometimes the epoch is not fully aligned with the interictal, preictal, or ictal period i.e. an epoch may have two types of period if an epoch is very large in size. Thus, it is important to extract global features from different epochs and local features within an epoch for correct seizure prediction. Moreover, the features extracted from spatially separated different channels of EEG signals should be incorporated to further improve the PA. In this paper, a novel approach is derived by exploiting spatiotemporal correlation of undulated global and local features within an EEG signal to find the transition of an event occurring during a seizure. Phase correlation [31] essentially provides relatively shifting information between current signals and reference signals of two correlated signals via Fourier Transformation. Thus, undulated global feature (UGF) can be determined using phase correlation. Paul et al. [31] demonstrated that the phase correlation is capable of detecting reliable motion between two images or blocks. In a similar fashion, the phase correlation extracting features between two adjacent epochs can capture the relative changes between two epochs of an EEG signal. This can be used to estimate the transition between interictal and preictal/ictal periods. However, sometimes this may not be adequate to identify the transition, if the transition is not aligned with the epochs. To avoid this situation, a local feature is also extracted from the signal fluctuation and deviation from the frequent oscillation within an epoch to achieve better accuracy and significant reduction in false alarms. Xie et al. [32] illustrated that fluctuation and deviation are able to identify defects of an image. This inspired us to use customized fluctuation and deviation [32] which can measure the fine changes of a specific epoch. Therefore, a cost function comprised of weighted fluctuation and deviation in each epoch is calculated in a temporal direction to extract the undulated local feature (ULF). Since EEG signals are non-stationary signals [33], the cost function of fluctuation and deviation (CFD) cannot fully identify the phase-lagging between two epochs alone. Thus, in this paper, both features (i.e. UGF and LGF) are used for advanced prediction of seizure onset with greater PA and low FPR. The paper is organized as follows: the data formation, the detailed proposed technique, feature extraction, classification, and post-processing are described in Section 2. Definition of SPH and statistical validation is described in section 3, and the detailed experimental results and discussions are explained in Section 4. Section 5 contains the analysis of results, and Section 6 the conclusion.