The international CHB-MIT database contains scalp EEG data collected at Boston Children's Hospital (CHB) from 23 pediatric patients with 9 to 40 recordings for each epileptic patient. This database was collected for epileptic seizure detection. One hundred and twenty-nine of those records that contain one or more seizures and 535 of the recordings contain no seizure activity. The recordings were sampled at 256 samples per second (256 Hz) with a resolution of 16 bits. Specifically, one hour for each patient selected to present an interictal period (seizure free) used for training data with 720 segments, and 40 minutes selected to be a preictal period with 480. All segment's length is 5 seconds. The seizure period is given in database description but its changes between one signal and another even in the same subject and it was segmented into 5-second windows.
The EEG is a non-stationary signal based on geometric characteristics that can be measured and detected using the shape of the corresponding element. Many reported methods are applied to solve the problem of automated peak detection in EEG signal data. A peak is defined as a local phenomenon, whereas a local peak may not be accepted as a true peak comparing with other peaks in the time series. A data point in a time series is considered as a local peak if : (1) it is a high and local maximum value in a window and it is not necessary that the value must be large or maximum in the time series; (2) not too many points have similar values in the window.
In this paper, a formal characterization of a peak in a time series is proposed to detect spike forms in order to predict the epileptic seizures in EEG signal. The proposed algorithm uses a raw time series data and does not require any pretreatment such as smoothing that eliminates some subjective aspects and details that are essential for the prediction process. To highlight the epileptiform in EEG signal, the selected element to extract the spike must be adapted to the geometric characteristics of the EEG signals. This section describes the operating process of the proposed algorithm that begins with searching for all maximum local peaks in a 5-second window. A segment can form a set of candidate peaks. Then each peak in the segment, if it occurs in a period less than 20 milliseconds and over 70 milliseconds and its amplitude should be 100 μV and more, is rejected[13–14]. Surviving peaks that exceed 70 milliseconds are discarded and considered as low frequency waves. The remaining peaks are accepted as spikes. In this work, a spike detection process is based on these studies for the extraction of epileptic spike.
The spikes detection process of EEG signal is represented by the flowchart in Fig. 1 and in the following steps:
1) For all the channels, EEG signal is separated into three periods, which are the interictal period chosen to be 1 hour, preictal and ictal periods, where the preictal time is 40 minutes after seizure onset[23–24]. For each period (interictal, preictal, and ictal), a sliding window is used to obtain segments with duration for 5 seconds to obtain the spike number in each segment.
2) Spikes are detected if the duration ranges from 20 milliseconds to 70 milliseconds and the minimum amplitude 100µV such as the sampling frequency of EEG signals is 256 Hz[13– 14,20]. EEG epileptic spikes can be extracted efficiently in the CHBMIT database.
Fig. 2 presents an example of three EEG signal periods: interictal, preictal, and ictal. One hundred segments in each period are randomly selected from the EEG CHB-MIT database. The length of each segment is 5 seconds. The variation of the spikes number distribution in the EEG periods is well distinguished where the spikes number of ictal segments is the maximum compared to the of the other two periods. The spikes number in preictal period is greater than interictal periods.
Table 1 represents the number of segments, the means of spike number in each period, and the maximum value of spikes/segment are 4, 7, and 11 in interictal, preictal, and ictal period, respectively. The comparison of the three EEG periods shows a significant differentiation according to their spike distributions.
EEG periods Segments number Mean Maximum spike/segment Interictal 100 0.43 4 Preictal 100 0.99 7 Ictal 100 3.13 11
Table 1. Spikes number in interictal, preictal, and ictal EEG periods
As shown in Fig. 2, the spikes number gradually increases as the seizure approaches and reaches a maximum of seizures. Based on this observation, the rate of peaks was chosen as a tool and as an indicator to predict epileptic seizures. Fig. 3 illustrates the proposed seizure prediction approach. CHB-MIT database (EEG scalp) was used to prove the robustness of the proposed method, and the seizure prediction process based on spikes rate is described in the following algorithm:
1) For all the channels, separate the EEG signal into interictal, preictal, and ictal periods. The preictal time is 40 minutes before the seizure event, and interictal period is chosen randomly from seizure-free recordings with a period of 1 hour.
2) After the extraction of the three period datasets (interictal, preictal, and ictal), a sliding window is used to separate each period in segment with 5 seconds.
3) Detect the spikes in each segment for all periods then calculate their total number of spikes.
j represents the period, i represents the number of segments in j period, and n(k) represents the number of spikes in the ith segment for the jth period in the above, and Nj is the total spike number of j period.
4) Determine the spikes number segments in each period and smooth Nj(i) using the average filter to obtain the new spikes number of i segment relative to its neighbors to equalize the spikes number in a period j. 7 is the length of average filter, and 1 is the moving step.
SN(i) is the smoothed spikes number, Nj(i) is the input segment, a is the neighbors of the ith segment, and M represents the length of the smoothing filter who's chosen to be 7.
5) The maximum value of the spikes number for the ith segment in interictal period is the threshold to predict seizure.
For each patient, Thresh represents the training data threshold. The alarm is triggered when the spikes value in preictal period exceeds threshold.
6) At least one alarm is triggered in the preictal period (during 40 minutes) in any channel, which may indicate an impending of epileptic seizure will occur in the near future.
The training data represented by 1 hour of interictal, 40 minutes preictal and 5 minutes seizure events were randomly selected from EEG data for each patient, and the test dataset presented by the rest of EEG data to evaluate the performance of the proposed method. Note that the threshold can be different for each patient.
Prediction of epileptic seizures and the triggering of an alarm or alert is the goal of different methods to have an effective algorithm to improve the quality of life of epilepsy patients by eliminating the damage caused by epileptic seizures that can be presented by loss of consciousness. The key point for successful prediction of seizures is to differentiate between preictal state and interictal state. Heretofore, tens of linear and nonlinear methods with univariate measures and multivariate measures have been partially successfully applied to predict epileptic seizure.
Before seizures event, the brain state is stable; there is no epileptiform discharge presented. With the discharge of some neurons that have gone to the abnormal state, the epileptiform, like spike, increases gradually. In Table 1, the statistical analysis of spike rates in different periods of EEG signal (interictal, preictal, and ictal) shows that there is a significant variation between them.
Each prediction method has been realized with a different anticipation strategy. Therefore, it can be concluded that no prediction time can be fixed and considered as a norm. Hence, it is possible to note that the prediction time varies from period to another and from records to another, even for the same subjects. The spike number can vary between different EEG signal periods, so it is necessary to have a specific setting parameter for each patient to make the spike rate threshold stability optimal for all recordings. The performance of the seizure prediction algorithm must be tested and evaluated on clinical cases and can be implemented on an epilepsy prediction device according to simplicity and understandable logic.
Table 2 and Fig. 7 present different proposed methods for the purpose to predict epileptic seizures; where each method is different from other methods according to the approach and strategy proposed. Some studies proposed to use epileptiform for seizure prediction with SPH 10 seconds. In the same estimated period, some studies obtained 10 seconds as prediction period; the research were based on Fast Fourier transform and Backpropagation Neural Network classifier, with classification accuracy rate of 89.67%, or the Wavelet Transform and Bayesian discriminant analysis was recently used, with classification accuracy rate of 93.62%. The change in the number of epileptiform changes significantly between the periods, which clearly allows anticipating the seizure state.
Ref. Database and methodology Prediction time Performance Li SF et al, 2013 - Freiburg database
- Low-pass filter, Morphology filter, Spike rate detection
10 seconds Sensitivity 75.8% Zandi et al, 2013 - Private data
- Histogram, Variational GMM, Zero crossing intervals
2 minutes Sensitivity 88.34% Zheng et al, 2014 - Freiburg database
- BEMD, Mean phase coherence
20 seconds Able to detect synchrony changes before the onset Zhang et al, 2014 - Freiburg database
- Higuchi FD, Bayesian LDA, Kalman filtering
2 minutes Sensitivity 89.33% Zhang et al, 2014 - Private data
- Approximate entropy
25 seconds Accuracy 94.59% Teixeira et al, 2014 - EPILEPSIAE database
- Auto-regressive modeling, Kruskal-Wallis test, ANN, SVM classifier
15.58 minutes Sensitivity 73.55, 24.83% Bandarabadi et al, 2015 - EPILEPSIAE database
- Amplitude distribution histograms, Spectral power features
8 seconds Able to predict seizures Bandarabadi et al, 2015 - EPILEPSIAE database
- Relative spectral power features, MRMR feature selection, Amplitude
distribution histogram, SVM classifier
5 seconds Sensitivity 75.8% Behnam et al, 2016 - CHB-MIT database
- Interpolated histogram feature, Seizure distribution model, Bayesian
classifier, Hunting search algorithm, MLP classifier
6.64 seconds Accuracy 86.56% Fujiwara et al, 2016 - Private data
- Time, frequency domain features, Multivariate statistical, process control
10 seconds Sensitivity 91% Fei et al, 2017 - Private data, CHB-MIT database
- Fractional Fourier transform, Modified LLE features, BPNN classifier
10 seconds Accuracy 89.67% Direito et al, 2017 - EPILEPSIAE database
- Auto-regressive modeling predictive error, decorrelation time, statistical
moments, energy, SVM classifier
5 seconds Sensitivity 38.47% Chu et al, 2017 - Private data, CHB-MIT database
- Spectral feature, Fourier coefficients
20 seconds Sensitivity 86.67% Zhang et al, 2018 - Private data
- A mathematical model
10 seconds Synaptic plasticity has influence on seizure period Yuan et al, 2018 - Freiburg database
- Wavelet transform, Diffusion distance, Bayesian discriminant analysis
10 seconds Sensitivity 93.62% Tsiouris et al, 2018 - CHB-MIT database
- Statistical features, Zero crossings,Wavelet transform, Power spectral,
Cross-correlation, Graph theory, LSTM
15 minutes to 20 minutes Sensitivity 99% Proposed method - CHB-MIT database
- Spike detection, average filter, threshold from training data.
- Alarm triggered
1 minute to
Able to predict seizure period with 92% for true prediction alarm
Table 2. Comparative study on recent works conducted on seizure prediction
Figure 7. Comparison of seizure prediction accuracy rate of the proposed method against the state-of-the-art methods.
In this work, an epileptic seizure prediction method is proposed based on the spike rate in the EEG signal. The algorithm detects spikes number in all channels over three EEG periods by applying the local maximum where the time range and amplitude of spike were given. The spike rate is smoothed with an average filter to balance the segments spike distribution in the same EEG periods, where the maximum number of spikes in the interictal period is used as a threshold and as index that there is an impending seizure in the near future. The alarm is triggered when the spike number in an interictal period segment exceeds the threshold. The CHB-MIT database is used to evaluate the algorithm, and it is shown that the spike rate increases with the occurrence of a seizure and reaches a maximum in seizure state. The proposed approach achieves a prediction rate up to 92% for all patients with at least one alarm is triggered at least in one channel. Comparing the epileptic seizure prediction methods such as correlation dimension, phase synchronization and other algorithms, the proposed algorithm allows to obtain a higher precision with a perfect prediction rate. In order to improve the quality of life of epileptic patients, after the validation and according to the simplicity of this algorithm, it is possible to make a portable device for monitoring epileptic seizures.
Epileptic seizure prediction based on EEG spikes detection of ictal-preictal states
- Received Date: 2019-07-01
- Accepted Date: 2019-12-23
- Rev Recd Date: 2019-10-16
- Available Online: 2020-02-17
- Publish Date: 2020-05-01
Abstract: Epileptic seizures are known for their unpredictable nature. However, recent research provides that the transition to seizure event is not random but the result of evidence accumulations. Therefore, a reliable method capable to detect these indications can predict seizures and improve the life quality of epileptic patients. Seizures periods are generally characterized by epileptiform discharges with different changes including spike rate variation according to the shapes, spikes, and the amplitude. In this study, spike rate is used as the indicator to anticipate seizures in electroencephalogram (EEG) signal. Spikes detection step is used in EEG signal during interictal, preictal, and ictal periods followed by a mean filter to smooth the spike number. The maximum spike rate in interictal periods is used as an indicator to predict seizures. When the spike number in the preictal period exceeds the threshold, an alarm is triggered. Using the CHB-MIT database, the proposed approach has ensured 92% accuracy in seizure prediction for all patients.
|Citation:||Itaf Ben Slimen, Larbi Boubchir, Hassene Seddik. Epileptic seizure prediction based on EEG spikes detection of ictal-preictal states[J]. The Journal of Biomedical Research, 2020, 34(3): 162-169. doi: 10.7555/JBR.34.20190097|