Prediction of Seizure in the EEG Signal with Time Aware Recurrent Neural Network

An epileptic seizure is a transitory incidence of symptoms or signs due to the extreme or irregular synchronous actions of neurons in the brain [1]. Generally, the occurrence of epilepsy in the brain is confirmed by the visual examination of long-term recorded scalp electroencephalograms (EEGs) and spots the incidence of epileptic seizure that utilizes a huge time to process [2]. Thus, the process of automatic diagnosis to identify the epileptic seizure could be importantly minimized the diagnosis duration. Numerous features are included for the automatic identification of seizures in the brain [3, 4].

The values used in the automatic identification of seizure are the connectivity of functional network properties, autocorrelation, nearest neighbor and likelihood synchronization, and EEG morphology [5-9]. The early detection of seizures is necessary to cure the disease. The recurrent features in the domain are identified through rhythmic actions, which are observed frequently in seizures [10]. The incidence of seizure can be identified from the features that are easily detectable namely principal components, spectral, statistical, and nonlinear features [11-14].

The above-mentioned features have shown excellence in the identification of certain kinds of seizures. The diverse nature of seizures made numerous complications progress a global feature for automation in seizure detection. Additionally, seizures are rarely occurring events and it is suitable in training complicated supervised learning processes of seizure with linear type classifiers, artificial neural networks, support vector machine, and other computational algorithms [15-17].

The earlier approaches lack accuracy in detection and also it takes a long time to process the data [18-21]. Currently, the Kraskov entropy with Least Square Support Vector Machine (LSSVM) method [22] has been developed for classifying seizure and seizure-free EEG signals, which support an automated prognosis of epilepsy. Also, a mixture of entropy with a Logarithmic Band Power (LBP) and K-Nearest Neighbor (KNN) classifier [23] has been presented for classifying EEG features into either epilepsy or autism spectrum disorder. Meanwhile, a Pairwise one-class SVM [24] has been applied for epilepsy detection and diagnosis. Though these methods were robust to categorize normal and seizure EEGs, the labeled training with the manual preparation is taking longer time and is labor intensive whereas these issues increase the data quantity. CNN correctly recognized irregular inter-ictal discharges as non-seizures, but could not detect the ictal state and slower oscillations. To improve the performance of CNN for detecting seizures' ictal state and slower oscillations, Recurrent Neural Network (RNN) is combined with the CNN model. Seizure detection at the early stage is necessary and it is significant to detect with computational algorithms.

Seizure detection is attained through the diverse mechanism and the drawbacks in the mechanism are rectified in the proposed Time Aware Convolutional Neural Network with Recurrent Neural Network (TA-CNN-RNN). TA-CNN extract the EEG signals with time context. The time and the feature integration in TA-CNN helps to accurately find seizure. TA-CNN not only capture dynamic changes of feature over time but also the seizer time context to detect. The time context is then effectively utlized in LSTM. For the investigation of the approach, fifty non-focal and focal signals were randomly selected from the publicly accessible EEG database. The proposed TA-CNN-RNN is compared with the existing algorithms Entropy+LSSVM [22], LBP-KNN [23], and P-one-class SVM [24]. The existing and proposed approaches are applied to the different variants of datasets namely CHB-MIT-EG, Bonn-iEEG, and VIRGO-EEG. The numerical outcomes of the experiment are evaluated using the performance metrics namely accuracy, precision, f-measure, and recall.

The remainder of the paper is emphasized as follows: previous works and literature is described in Section 2, the detection and diagnosis of epilepsy in the EEG signal are attained by the proposed deep learning approach is detailed in Section 3, the numerical outcome of the experiment is given in Section 4 and the proposed deep learning approach is concluded with a future suggestion.

In a deep learning-based system, EEG signal pre-processing commonly involves three steps: noise removal, normalization, and signal preparation for deep learning network applications. In the noise removal step, finite impulse response (FIR) or infinite impulse response (IIR) filters are usually used to eliminate extra signal noise. Normalization is then performed using various schemes such as the z-score technique. Finally, different time domain, frequency, and time-frequency methods are employed to prepare the signals for the deployment of deep networks. In this research work, Time-aware CNN (TA-CNN) is proposed which incorporates position information into CNN through an attention mechanism.

The suggested model accurately represents the time series dynamics present in EEG data. The temporal patterns of the signal vary greatly between the two time intervals, and even among the several time steps that make up the same interval. Therefore, the associated patterns require modelling at various time scales. So that the model may prioritise different scales at different time steps, we devised an efficient mechanism called TA-CNN that makes use of temporal context information to adjust the features of different scales. The proposed TA-CNN can capture important features for relation extraction effectively.  To obtain more precise and fast detection of epileptic seizures TA-CNN is combined with RNN to correctly detect the ictal state and slower oscillations part of the EEG signal. RNN can learn temporal and context features, especially long-term dependency between features with a seizure while TA-CNN is capable of catching more capable features with positional information. Therefore, the combination of TA-CNN and RNN becomes necessary for a higher-quality epilepsy classification.

3.1 Signal pre-processing

In Figure 1, the overall illustration is the block diagram of the proposed TA-CNN-RNN. The signal pre-processing is intended to acquire the significant features with a greater probability of pre-ictal portion of correlation, the process of pre-processing is attained to the data before making the training and testing process with TA-CNN-RNN. The structure of TA-CNN-RNN is shown in Figure 2.

1.png

Figure 1. Overall process of epilepsy classification

2.png

Figure 2. Structure of TA-CNN-RNN framework

3.1.1 Noise removal

The EEG’s high quality is greatly influenced by the noise which is a huge enemy of the EEG signal. The incidence of noise is decisive for the accurate investigation of EEG. The incidence of noise in the EEG signal is removed by the Least Mean Square Adaptive Noise Cancellation (ANC). The unwanted information in the power line of the EEG signal is removed successfully by the ANC scheme.

3.1.2 Normalization

Generally, normalization is utilized in carrying the two signals to a similar or predefined series. A distinctive example of a predefined series is the statistical discernment of the normalization that is converting the signal where the mean value is 0 and the standard deviation value is 1. In this proposed scheme, the z-score technique is utilized for the normalization and this technique unveils the performance by flattening the signal. The value of the z-score is equated in Eq. (1):

$z- score =\frac{ { score }- { mean }}{ { standard\,\, deviation }}$            (1)

This normalization preserves the correlation between the normalized and original EEG signal whereas it considerably minimizes the selection bias. After completing the normalization process deep learning approaches are incorporated and the normalization makes the learning approach easier.

3.2 Significant feature extraction and signal preparation for the categorization of signal

The raw EEG signal is suspicious for redundancy and hence it is necessary to mine the explanatory parameters. The significant tool for investigating the non-stationary signal is accomplished by the Time Aware Convolutional Neural Network. Diverse time fields, frequencies, and time-frequency are incorporated to formulate the signals for the disposition of the deep neural network. The information about the location is taken into the CNN via an attention mechanism. The proposed TA-CNN can acquire the significant features in the extraction relation proficiently. The time intervals are deployed to acquire the short and long-term interest. The TA-CNN is utilized for the feature selection and suspected portions in the signal are acquired. Every attribute is a rate of discrete coincidence and utilizes the data among the measurement of similarity in the attributes S(a, b) as:

$S(a, b)=\sum_{b \in B} \sum_{a \in A}(a, b) \log \left(\frac{p(a, b)}{p_1(a) p_2(b)}\right)$                     (2)

In Eq. (2), p(a, b) is the probability of combined a and b attributes in the function distribution. The p₁(a) and p₂(b) illustrated the marginal probability. The combination of probability in the equation is acquired from the neural network. The proposed scheme identifies the significant features that signify the component's magnitude and every signal represents a diverse spectrum of components in original data that is adequate and informative for diverse EEG patterns. The acquired features are fed to the classification and diagnosing part.

The Long Short Term Memory network is a special kind of recurrent neural network and it is an order sequence of data. The main intent of the proposed scheme is to classify the sequence of data among the pre-ictal and inter-ictal states that indicate the probability range of high or low in every class. This process is attained by the Recurrent Neural Network (RNN) to accept the sequence of temporal information and preserve the appropriate information. The network is formulated to estimate the data batches that create an output. It is elected to utilize the data with high-size processed data. This approach has three networks and one output layer.

The acquired features are passed to the first two layers whereas the values are elected randomly that minimize the error. This process returns the output after processing the data in the preceding layers and probability is estimated by the softmax activation function. The RNN selects the data that is remembered or forgotten. In this study, the LSTM [28] is adopted as the RNN, which is made up of a cell and 3 gates (input, forget, and output). At time t_i, the learning network is updated from Eqns. (3-8):

$i_{t i}=\sigma\left(W_{a i} a_{t i}+W_{h i} h_{t i-1}+y_i\right)$                  (3)

$f_{t i}=\sigma\left(W_{a f} a_{t i}+W_{h f} h_{t i-1}+y_f\right)$                  (4)

$g_{t i}=\tanh \left(W_{a c} a_{t i}+W_{h c} h_{t i-1}+y_c\right)$                 (5)

$c_{t i}=i_{t i} \odot g_{t i}+f_{t i} \odot c_{t i-1}$                   (6)

$o_{t i}=\sigma\left(W_{a o} a_{t i}+W_{h i} h_{t o-1}+y_o\right)$                 (7)

$h_{t i}=o_{t i} \odot \tanh \left(c_{t i}\right)$                     (8)

In Eqns. (3) – (8), i_ti is the input gate, f_ti is the forget gate, g_ti is the memory cell candidate, c_ti is the memory gate, o_ti is the output gate, h_ti is the hidden states in the LSTM, σ is the activation function, $\odot$ is the element-wise multiplication, a_ti is the input vector value at the time ti, W_ai, W_af, W_ac, W_ao are the weight matrix of a_ti’s input, forget, memory, and output gates, respectively. Also, W_hi, W_hf, W_hc, W_ho are the weight matrix of h_ti’s input, forget, memory, and output gates, respectively. As well, y_i, y_f, y_c, y_o are the bias offset of every gate.

In the RNN, the information is propagated from the front layer to the back layer and the drawback in the one-way approach in the learning scheme is rectified by this approach. The RNN has an inverted and positive LSTM that captures the significant feature and also combines the features in the forward and backward by utilizing the element-wise sum equated in Eq. (9):

$h_i=\overrightarrow{h_l} \odot \overleftarrow{h_l}$                     (9)

The main intention of the attention strategy is identified and the input information is significant in the process of training where the attention is close to the data. From the RNN layer the h_i and TI is assigned to the matrix M.

$M=\left[m_1, m_2, \ldots \ldots, m_L\right]$                   (10)

In Eq. (10), the length is signified as L of the input signal and the M_tr is acquired by the weighted alpha for every m_i as shown in Eqns. (11) - (13):

$M_{t r}=\tanh (M)$                 (11)

$\alpha=\operatorname{softmax}\left(\omega^L M_{t r}\right)$                   (12)

$M_{t r}=M \odot \alpha^L$                 (13)

The dimension unit of hidden value in the learning unit is given as M belongs to d(h*L), the parameter of trained vector and transpose vector are ω and ω^L respectively. The signals are acquired that is assigned as d_h*L. The process of convolution is carried by the kernel W:

$A_j=f\left(X_j \odot \mathrm{W}+\mathrm{b}\right)$                (14)

In Eq. (14), the convolution operator is signified as $\odot $, the effect of bias is signified as b, the activation function for non-linear data is signified as f and the ReLU activation is equated in Eq. (15):

$f(a)=\max (0.1 a, a)$                    (15)

The dropout layer in the learning process is equated as:

$h_n=\omega_n^L(r \odot b)+b_n$                   (16)

In Eq. (16) the same signal shared among the values is signified as r and each element of r has the probability value 0 for p and 1 for 1-p. The dropout layer is assigned with the subsequent layers and the rates of probability are 0.3, 0.3, 0.3 and 0.5.

The overfitting issue in the signal processing is rectified by the enhanced generalization capability and the cost function for the regulation is added as:

$J(\theta)=\sum_{i=1}^m t i_i \log \left(b_i\right)+\lambda\|\boldsymbol{\theta}\|_F^2$                  (17)

In Eq. (17), the hyperparameter utilized for regularization is given as λ.

The RNN attains the classification that relies on the sequence of signals that are analyzed and the categorical cross-entropy is utilized as a loss function. The overlay performance in the RNN in the input segment of the initial layer is altered to acquire the batches of signals whereas the length of the signal is longer in the process of training and simple in the testing. Additionally, the internal values are estimated in the preceding batch that is no longer to pass every substantial batch of data due to alterations in the architecture of the network to acquire the efficient signal batches count. The utilized training epoch in the RNN considerably minimizes the incidence of error during the process of training which eventually increases the rate of accuracy.