Preictal and Interictal Recognition for Epileptic Seizure Prediction Using Pre-trained 2D-CNN Models

In this study, a deep learning method for preictal/interictal recognition, which is an important step in seizure prediction applications, is proposed. Feature extraction was employed for differentiating the preictal and interictal states from scalp EEG signals using trained CNN models with millions of images. EEG signals were converted to spectrogram images and were used as input data to 2D CNN models. The preictal/interictal differentiation was performed using the extracted features. The flow diagram of the study is given in Figure 1.

1.png

Figure 1. Flow diagram of the proposed method

2.png

Figure 2. 18-channel samples and the electrode positions placed according to the 10-20 system during EEG recordings

2.1 EEG dataset

In this study, the publicly available CHB-MIT scalp EEG dataset was used [23]. EEG signals were collected from children who were receiving treatment for seizures in Boston Children’s Hospital. All EEG signals were obtained from 18 to 23 channels and were sampled at 256 Hz. Figure 2 shows the 10-20 System of EEG Electrode Placement used for collecting EEG signals.

Total duration of the EEG recordings was 983 hours. Most EEG recordings consisted of 1-h epochs, while others were 2 to 4-h epochs. The seizure onset and offset times were marked by a clinical expert. In the annotation file of the data set, channel changes, seizure start/end information was given. Table 1 shows the information for the CHB-MIT EEG data set.

Table 1. CHB-MIT EEG data set description

For this study, the preictal interval was determined as the 30-minute interval ending 30 minutes before the seizure onset. This duration was determined since it is highly important to give an effective time for the physician to intervene in the patient before the seizure onset [16]. Moreover, the preictal interval is an arbitrary choice of investigators since there is no precise time limit for the preictal state [2]. Therefore, in this study, the preictal window of 30 min was chosen. The 10-min time period after the seizure was determined as the postictal state [24]. There were various seizures that occurred over a short interval in the CHB-MIT EEG data set, of which the seizure that occurred first was included in the analysis There were 20 cases that fulfilled the inclusion criteria (see Table 1). The time periods other than the preictal, ictal, and postictal periods were accepted as interictal, which occurred at least two hours away from the seizure [2]. In addition, segments were chosen as non-overlapping 5 sec as in the study [2]. Figure 3 shows an example of a signal for the identification of preictal and interictal periods.

3.png

Figure 3. Brain activities of epilepsy patients include interictal, preictal, ictal, and postictal states. The first 30-min period before each seizure was defined as the preictal period

2.2 Preictal state detection

The preictal detection approach consisted of the segmentation of EEG signals, conversion of each segment into spectrogram, extraction of deep features from spectrogram images, and the classification stages (Figure 1).

2.2.1 Segmentation

For one patient who had a history of five seizures, a 150-min (5x30) preictal state and a 150-min interictal state were randomly selected. The time periods defined as ictal in the annotation file were not used. The EEG signals were then segmented for each case in the data set. Figure 4 shows some examples of preictal and interictal signals that were divided into 5-sec segments. The sliding window method was used for segmentation, which can be used with or without overlap. The overlapping sliding window analysis is highly useful for data continuity although it causes redundant information. The analysis of 18-channel data of each patient resulted in excessive information density, which can cause serious problems both in terms of processing load and timing. Accordingly, the EEG signals were divided into non-overlapping segments.

2.2.2 Spectrogram

Since a 2D CNN model was used in the study, the segmented raw EEG signals were converted to 2D image format. Methods such as Short-time Fourier Transform (STFT) and Wavelet transform were used effectively in image conversion and feature extraction since these methods retain the time-frequency information in the analysis of signals [25]. All preictal and interictal activities were divided into 5-sec segments. While obtaining the spectrogram, the hamming window width was selected as 16 ms, overlap 8 ms, and 512-point Fast Fourier Transform (FFT) was used. The Viridis scales were employed to provide a color map. Figure 4 shows the spectrograms examples of preictal and interictal segments. The obtained spectrogram images were of 875 x 656 pixels. The pre-trained VGG19, ResNet and DesNet models have an original input size of 224 x 224 pixels. Additionally, using large images for input requires more hardware resources and increases processing load. Therefore, each spectrogram image was resized to 224 x 224 pixels.

4.png

Figure 4. 5-sec segmented preictal/interictal EEG signals

2.2.3 CNN

CNN models have been shown to provide successful results in classification and detection in medical applications [14, 26]. These models became more popular with the success in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012 by Krizhevsky et al. [27]. Moreover, the advent of VGGNet and ResNet further increased this popularity in 2014 and in 2015, respectively. A good training process is needed for a CNN model to achieve a good classification or recognition. Additionally, a large amount of data is required for good training, which mostly cannot be obtained. In such cases, using pre-trained models with large data sets produces more successful results. Instead of re-training a CNN model, it is easier to perform a process using the weights of pre-trained models. This method is called Transfer Learning, which allows a more efficient feature extraction from smaller data sets and also allows the training of small data sets with low computation costs [28]. In this study, VGG19, ResNet, and DenseNet models were used to extract deep features. For each spectrogram image, VGG19, ResNet, and DenseNet produced feature vectors of length 4096, 2048, and 1920, respectively.

VGG19: CNN models, which began with LeNet and then become popular with AlexNet, were effectively used by the Visual Geometry Group of Oxford University (VGG) in the development of VGG16. VGG16 consists of 13 convulsion layers and 3 fully connected layers and VGG19 consists of 16 convulsion layers and 3 fully connected layers. There are also 5 max-pooling and softmax layers which are the final layers in both models. While softmax is used to classify data, ReLu is used as the activation function [19, 29, 30].

ResNet: The earliest CNN models had a remarkably low number of layers. Of note, VGG16 had 8 layers and VGG19 had 16 layers. However, with the rapid developments in GPU technology, the number of layers in CNN models increased. One of these models is ResNet architecture, in which different architectural structures with 50, 101, and 152 layers have been used. In deep CNN architectures, the number of layers was increased, which made the training difficult. As a result, the input and gradient values started to vanish. In order to solve this problem, the skip connection technique was used in ResNet, whereby the output of one layer was added to the next input by skipping certain layers and thus a more effective training model was provided [31, 32].

DenseNet: DenseNet has a structure in which the previous layer properties are forwarded to all subsequent layers. Unlike ResNet, DenseNet transfers the previous layer information to all subsequent layers, instead of skipping some layers and transferring the previous layer information to the next layers. This provides a stronger flow of information between the layers. Another advantage of DenseNet is that it creates inputs by adding information from the previous layer to the next layer, rather than collecting features when creating layer inputs [33].

2.2.4 Support Vector Machine

Support Vector Machine (SVM) is a machine learning algorithm used for separating data sets with two or more classes by finding an optimal hyperplane. SVM reduces structural risks and uses support vectors to separate classes. After determining support vectors, the most appropriate hyperplane to separate the data set is found. In the two-class data set $\left\{x_{i}, y_{i}\right\} i=1,2,3,4, \ldots, k, y_{i} \in\{-1,+1\}$ are the class labels in the dataset [34, 35]. Inequalities that define the most appropriate hyperplane to separate classes are described as follows:

$\begin{array}{cl}w \cdot x_{i}+b \leq 1, & y=-1 \\ w. x_{i}+b \geq 1, & y=1\end{array}$       (1)

As a result, the obtained separation function for the two-class data set that can be separated linearly is as follows:

$f(x)=\operatorname{sign}\left(\sum_{i=1}^{k} \alpha_{i} y_{i} K\left(x. x_{i}\right)+b\right)$             (2)

Here, α represents the Lagrange multiplier, b represents bias, $x_{i}$ represents support vectors and K represents kernel functions [34, 35]. The Kernel functions used in the study included Radial Basis Function (RBF), Polynomial, and Linear.

2.2.5 Performance evaluation

The k-fold cross-validation method was used to evaluate the performance of the proposed method. The k value was determined as 5 and thus the data set was divided into 5 parts. Four parts were used for training and the remaining part was used for testing. This procedure was applied to all parts. The average of five values was calculated for performance evaluation. The parameters used for performance comparison were defined as True Positive (TP), which represented the number of precisely defined preictal segments, as False Negative (FN), which represented the number of incorrectly defined preictal segments, as True Negative (TN), which represented the number of correctly defined interictal segments, and as False Positive (FP), which represented the number of incorrectly defined interictal segments.

Sensitivity $=\mathrm{TP} /(\mathrm{TP}+\mathrm{FN}) \times 100$                 (3)

Specificity $=\mathrm{TN} /(\mathrm{TN}+\mathrm{FP}) \times 100$                 (4)

Accuracy $=(\mathrm{TP}+\mathrm{TN}) /(\mathrm{TP}+\mathrm{FP}+\mathrm{TN}+\mathrm{FN}) \times 100$                  (5)

In the proposed model, EEG signals were transformed into a spectrogram image using MATLAB software. Other applications were implemented using the KERAS 2.1.6 library in Python 3.5. Training and testing of the model were performed on a Linux server with NVIDIA GTX 1080 graphics card.

Ictal segments are taken into consideration in seizure detection studies, whereas these segments are ignored and a seizure is tried to be determined by comparing preictal and interictal segments in seizure prediction studies. There is no clear procedure to determine preictal activities [2]. The general view is that the preictal interval covers a wide period of time which could be as long as two hours before the onset of a seizure [2, 24, 37]. Accordingly, in the current study, the 30-min interval ending before the onset of seizure was selected as the preictal period. It is important to predict a seizure early enough to allow a clinical intervention [16], which could make the disease more controllable [2].

To date, there have been numerous studies investigating the discrimination of preictal and interictal state. Song et al. proposed a method using the Freiburg data set extracted from 6-channel iEEG recordings of 21 patients. The authors divided the EEG data into parts using 5-second non-overlapping sliding windows from 30-min preictal segments. Using the Sample Entropy properties of each segment, the data was classified as preictal and interictal with 86.75% sensitivity and 83.50% specificity [38].

Lin et al. presented a global feature extraction method from artifact-free EEG segments of 8-channel sEEG data of five patients. Preictal state was selected as the 3-minute section before a seizure onset and the interictal state was selected as the section one hour away from the seizure. The authors used 20-sec overlapping and 30-sec sliding windows and extracted 216 features from each segment. Subsequently, five properties were selected for each patient and the preictal and interictal states were classified with 97.50% accuracy, 96.92% sensitivity, and 97.78% specificity [37].

Cho et al. applied their method for 5-min sEEG recordings of 21 cases in the CHB-MIT dataset. The preictal state was determined as the section 5 minutes before the onset of a seizure and the interictal state was determined as the section 30 minutes away from the seizure. Features were extracted from EEG signals using one-second sliding windows with a band-pass filter, empirical mode decomposition, and filtering algorithms, with an average accuracy of 83.17% [8].

Zhou et al. obtained the iEEG recordings of 21 cases from the Freiburg data set and the sEEG recordings of 23 cases from the CHB-MIT dataset. Both data sets were segmented using a one-second sliding window and were used as input to the 3-layer CNN. EEG data were classified into three different types as preictal/interictal, interictal/ictal and preictal/interictal/ictal. However, the study did not include any information on preictal/interictal selection and a mean accuracy of 95.6%, a mean sensitivity of 94.2%, and mean specificity of 96.6% were achieved in the CHB-MIT dataset [39].

In the method proposed by Elie et al., bispectral properties obtained from 16 channels using iEEG recordings of 3 dogs from the Neurovista data set were used as input to 5-layer ANN. The preictal part consisted of a 5-minute section one hour before the seizure onset, and the interictal state consisted of a 5-minute section four hours away from the seizure. The authors used 30-sec non-overlapping segments for feature extraction and reported that the best testing accuracy obtained for preictal and interictal classification was 78.11% [40].

Tsiouris et al. performed the preictal and interictal differentiation using local and global characteristics from sEEG recording of 24 cases in the CHB-MIT dataset. The preictal state was determined as the section two hours before the seizure onset and the remaining part were determined as the interictal state. The model achieved a sensitivity of 85.75% and a specificity of 85.75% in the classification of preictal and interictal states [2].

Table 6 shows a comparison between the proposed method and those proposed in previously published studies. As shown in Table 6, previous studies determined the preictal interval as the section immediately before the onset of the seizure. In the method proposed in the present study, however, the 30-min interval ending 30 min before the onset of the seizure was selected as the preictal state. Our goal in identifying preictal activities was to examine the predictability of a seizure that could occur at least 30 minutes before the seizure. Another reason was that predicting a seizure 30 minutes earlier is highly important for prompt control of the disease. Our results showed that the preictal region can extend up to one hour before seizure onset. Furthermore, the classification of preictal and interictal states with an accuracy of over 90% showed that the duration of the preictal period showed a wide variation, as reported in previous studies.

Table 6. Comparison between the proposed method and the methods proposed in previous studies

** Time before a seizure onset

*** Time set to end 30 minutes before a seizure onset

The advantage of using pre-trained CNN models was that it produced effective features since they were trained on large data set. In this way, more successful classification results can be achieved by using a limited number of training and test data. In doing so, the difficulties in adjusting the hyperparameters in a CNN model design using one-dimensional EEG signals were eliminated. In traditional feature extraction approaches, the extracted features have a significant effect on the classification performance of the system. In these systems, not only the feature extraction process but also the experience of the operator affects the classification performance. Meaningfully, the proposed CNN model was remarkably successful since it extracted features from the data set itself.

The limitation of the study was the prolonged duration of the preictal/interictal interval examined.