Accurate and Efficient Differentiation Between Normal and Epileptic Seizure of Eyes Using 13 Layer Convolution Neural Network

Epilepsy is considered to be one of the widely recognized neurological illnesses portrayed by epileptic seizures and World Health Organization (WHO) indicates it to be the second crucial neurological ailment apart from stroke [1]. The occurrence of seizures might not depend on circumstances or host attributes [2]. The ill effects are experienced by the epilepsy patients in the course they can't secure themselves and are prone against suffocation, demise, or injury due to due to fainting and traffic accidents [3]. Until this point in time, this ailment is essentially given with prescriptions and medical procedure; there is no complete recovery and no effectual medicines with anticonvulsants for all epilepsy kinds [4].

 Electrooculography (EOG) is adopted for identifying epilepsy as the new area of research. Epilepsy in previous research works were detected using EEG signal which is obtained from subjects scalp. The voltage produced in the scalp is taken by using different electrodes positioned on the scalp [5]. An immediate assessment by a doctor is necessitated for recognizing EEG as a significant measure of time and exertion [6]. Moreover, specialists with contrasting degrees of analytic experience here and there report discrepant conclusions on the symptomatic outcomes. In this manner, the advancement of a computerized, PC supported strategy for the analysis of epilepsy is direly required [7].

The potential difference between cornea and retina voltages obtained using electrodes placed on lobes of the subjects [8] is estimated by Electrooculography (EOG) approach. The subsequent sign is known as the electrooculogram and its applications lies in ophthalmological verdict and in recording eye movements. EOG doesn't gauge reaction to individual visual improvements in contrast to the electroretinogram [9]. A sets of terminals are ordinarily positioned either above and underneath the eye or to one side and right of the eye for quantifying eye development. On the off chance that the eye moves from focus position toward one of the two electrodes, this electrodeode "sees" the positive side of the retina and the contrary terminal "sees" the negative side of the retina [10] which leads to a electrodes potential difference. The recorded potential is a proportion of the eye's position, supposing that the resting potential is steady.

EOG signals are greatly utilized for automatic seizure recognition by employing various methodologies [11]. Automatic seizure detection systems’ performance mainly relies on EOGs behaviour by means of extracting features. Various researches are carried out for feature extraction and selection techniques [12] and these are carried out in time-domain, frequency-domain or the integration of both the domains. There involve three main challenges in the above domains based approach. The first most challenge is that, very sensitive and robust scarce) to acute discrepancies in seizure patterns which is due to the non-stationary EEG data stationary in addition modifications in its statistical features through different subjects and over time for the same subject [13]. The various artifacts such as muscle activities, eye-blinks, and environmental white noise influences on EEG data acquisition systems which is the second case. The seizure detection systems performance accuracy is extremely affected due to the various noise sources which may modify the EEG features.

A Noise aware Epileptic Seizure Detection using Thirteen Layer Convolution Neural Network (NESD-TLCNN) is introduce to mitigate these issues. It also concentrates on prediction and precise determination of discrepancy among the normal eyes and the eyes with epileptic seizure regardless of noise presence. Various procedures are involved in this work for accurate noise concerned epileptic seizure detection.

1.png

3.1 Noise estimation and removal using hybrid dynamic time wrapping based hidden Markov model

Hidden Markov Model has employed to deal with the estimation of reference patterns, besides the generalization of training data with the reference patterns have done through Dynamic Time Wrapping, which previously determines the distance within recorded speech and a template. Then, the distance of the signals has estimated at every instant that accompanies the warping function. Additionally, the clusters have been trained by HMM that passes within clusters according to the probabilities provided by several models. Eventually, the recoded voice has attained nearly appropriate segmentation by this algorithm. The recoding has processed at noisy places. In some instances, the segmentation process might face the challenge of acquiring the inconsistent values from the algorithm, as it affected by the background noise. Due to the noise interruption, it necessitates increasing the cut off for silence, since it could not expect to be zero. The process of pre-processing using Hybrid Dynamic Time wrapping using HMM is shown in Figure 1.

A method of on-time optimal alignment for the test and the reference patterns has provided by the DTW algorithm in order to furnish the average distance corresponding to the optimal warping path. DTW is known to be a pattern matching algorithm that relies on Bellman’s principle of optimality with a non-linear time normalization effect.

The two sequences of the feature vector have aligned with the alignment procedure generated by DTW. i.e. T₁, T₂, …, T_n and R₁, R₂, …, R_n.

Here, a function f(i,j) signifies a distance between two feature vectors Ti and Ri that has considered as local distance. The local distances have iterated and added them for the optimal predecessor to calculate the global distance. The following equation expresses the optimal global distance:

D(i,j) = min _{m<=I, k<=j} [D(m,K)]+f(i,j)    (1)

In addition, DTW enables the establishment of time scale alignment within two feature vectors, besides it has the capability to procure the long-range dependencies in acoustic data and has the proficiency to attune with the variations in eye movement. The aforementioned factors reveals that the DTW is capable of performing the wide-ranging observation, and HMM is efficient enough to solve the assessment of the detail. So the conclusion can be focused on an amalgamation of DTW and HMM approaches as regards the design of effective framework. The following features can be considered to build the advanced DTW:

The provided input feature vector sequence T₁, T₂,…,T_n and a sequence of states R₁, R₂,..,R_n (DTW template) can be enforced with the method of DTW. Each states in the advanced DTW has described by the set of parameters.

1. The set of model parameters has represented by π

2. The estimation of the quality of the model has formulated as,

$f(y \mid \pi)=\sum_{q \in Q} f(y, q \mid \pi)$    (2)

where, the set of overall valid paths has indicated by Q.

3. The quality of path has estimated by,

$f(y, q \mid \pi)=f(q \mid \pi) f(y \mid q, \pi)$    (3)

in which, $\mathrm{f}(\mathrm{q} \mid \pi)$ signifies a prior measure for traversing through path q, whereas $\mathrm{f}(\mathrm{y}, \mathrm{q} \mid \pi)$ indicates a posterior quality measure.

The Prior path traversal measure has expressed as,

$f(q \mid \pi)=\left[\sum_{i=1}^{L(q)} a_{q_{i}}^{W\left(q_{i}\right)}\right]^{p(q)}$    (4)

Here, L(q) stands for the sequence of the connected segments q₁, q₂,…,q_n of valid path, while W(q_i) denotes the weight of the ith segment q.

$W(q)=\sum_{j=1}^{r_{q}} w_{q} . j$    (5)

in which, the number of grid points denoted by r_q. As a path independent, p(q) has represented as ρ. The estimation of the posterior path quality has defined as the entirety of segment quality that make path.

$f(y \mid q, \pi)=\prod_{i=1}^{L(q)} b\left(v_{q i} \mid \pi\right)$    (6)

Here, the quality measure of the segment $\gamma$ and its respective grid points has symbolized by $\mathrm{b}\left(\mathrm{v}_{\mathrm{qi}} \mid \pi\right)$.

$b\left(v_{q i} \mid \pi\right)=\prod_{j=1}^{L(q)} b_{q}\left(v_{q}(j) \pi\right)$    (7)

Here, the weighted and the estimation of normalized local similarity for grid point has notated by $b_{q}(v \mid \pi)=[p(v \mid \pi)]^{p(q) w_{q}(v)}$, whereas the Gaussian probability density function has represented by $p(v \mid \pi)$. The proposed method is not functioning in a statistic way, as the measurement of path quality might rely on the path and weighting, besides the normalization process disrupt the scenario of statistical feature.

The forward and backward algorithms has applied to determine the path quality, in which forward variable α(u) has designated as the total of the path quality measure across overall valid path from beginning to ending.

$\alpha(u) \triangleq \max _{q \in Q, \text { start }, u}\left\{\prod_{i=1}^{L(q)} a_{q_{i}}^{\rho W\left(q_{i}\right)} . b\left(V_{q_{i}} \mid \pi\right)\right\}$    (8)

Here, the set of parameters π has determined and V = V¹, V²,....,V^(n-1), Vⁿ presented as training patterns, with the intention of augmenting f(V/ π) as follows,

$\hat{\pi}=\arg \max _{\pi} f(V \mid \pi)$    (9)

3.2 Eye feature extraction

EOG signal characteristics are defined by measuring the variation in amplitude, in volts, with respect to time. The signal has a repetitive pattern for various activities. The development of the classifier and its performance depends on the selection of the right extracted features. Therefore, feature extraction plays a significant role in achieving better accuracy from the classifier.

A feature is defined as a structural, specific or characteristic measurement of the segment, which is extracted from certain pattern. The extracted features can be used to classify the signal into groups by defining their key factors. The process of feature extraction from the EOG signal and optimization of the extracted features are detailed in this section.

It is mandatory to identify the various forms of eye movement to utilize the analysis of epileptic seizure for recognizing the activity. There are three simple kinds of eye movement have identified which has eased by EOG: Saccades, Fixations, and Blinks.

3.2.1 Saccades

By nature, the eyes do not being static at the time of seeing an object optically, rather than they keep moving to construct an outline through the appealing portions of that object, only because of the fovea, a tiny pit located in the centre of the retina that enables the vision sharper. The concurrent activities of both eyes has termed as Saccade. The saccadic period relies on their saccade amplitude that is known as the angular distance travelled by the eyes at the time of this activity. The saccadic eye movements usually occur with the amplitude of 20 degrees, and 10 to 100 ms of duration.

3.2.2 Fixations

Fixations refers to the fixed positions of the eyes when the focus in the visual object has retained at a single spot. In general, fixations have represented as the duration within every two saccades. The typical duration of fixation set within 100 ms to 200 ms.

3.2.3 Blinks

The Precorneal Tear Film is known to be the front portion of the cornea which has surrounded by a thin fluid that spreads over the region of the cornea through the action of open and close made by the eyelids, namely blinking. Generally, the average rate of blink rests within 12 and 19/min at leisure times. However, there is a number of ecological factors, like brightness, humidity, and temperature, as well as the physical exertion due to mental workload/extreme tiredness that impact the blink rate, where the average blink raises amid 100 and 400.

3.3 Optimal feature selection using particle swarm optimization

Pre-classification process, the feature selection has considered to be a primary phase for eradicating the redundant features, for choosing the parsimonious, information-rich features, and for evading the outfitting at the process of classification. The algorithms of feature transformation and feature selection have widely utilized for diminishing the dimension of feature and choosing valuable features. This study tends to choose the information-rich weighted features by proposing the PSO based feature selection method. In a traditional way, the particles have arbitrarily initialized. Nevertheless, this work utilizes the mixed initialization methodology, in which the initialization of 50% particles includes a lesser amount of features (10% of overall features), and the rest of the particles utilize a huge amount of features (60% of overall features).

The evaluation of fitness has viewed as the primary phase in the PSO based feature selection. In general, the design of fitness function significantly includes two standard metrics, i.e. the accuracy and error rate of classification. Nevertheless, such metrics would not be adequate to assess the eminence of the particles while working with the imbalanced dataset, since they delude the efficiency of classification because of the focus on the impact of majority class. So that, this research tends to determine the fitness of each particle by designing the novel fitness function, in which the Geometric mean (G-mean) has exploited to assess the overall performance of the classification.

Fitness $=\alpha *(1-$ Gmean $)+(1-\alpha) *\left(\frac{\text { number of selected features }}{\text { All features }}\right)$    (10)

Here, the relevant significance of the classification performance (G-mean) has represented by α, whereas the relevant significance of the number of features has indicated by (1-α). According to compare with the number of features, the requirement of classification performance is high, so the value for α defined as 0.8. The quality of every particle has estimated based on a fitness function. Then, the algorithm upgrades the pbest and gbest subsequently the velocity and position of each particle have updated by the algorithm. Two scenarios decide the update of current pbest and gbest, i.e. i) if the classification performance (G-mean) of the particle’s new position was superior to the former pbest and the number of features was lesser than earlier pbest. ii) if the number of features was lesser than former pbest and the classification performance (G-mean) of the new position was identical/better than the current pbest. The updating of the gbest follows the same procedure. The position of a particle signifies the selected feature subset. V-shaped transfer function has enforced inn the suggested binary PSO, in terms of converting the velocity from constant space to probability space, as follows:

$S\left(x_{i}^{k}(t)\right)=\frac{2}{\pi} \arctan \left(\frac{\pi}{2} x_{i}^{k}(t)\right)$    (11)

$\begin{aligned} v_{i}^{k}(t)=w * v_{i}^{i} &(t)+c_{1} * \text { rand } \\ & *\left(p b e s t_{i}^{k}(t)-x_{i}^{k}(t)\right)+c_{2} \\ & * \text { rand } *\left(\operatorname{gbest}_{i}^{k}(t)-x_{i}^{k}(t)\right) \end{aligned}$    (12)

The subsequent rule of position update must be applied, since the v-shaped transfer function has been used.

$x_{i}^{k}(t+1)=\left\{\begin{array}{l}\tilde{x}_{i}^{k}(t) \text { rand }<T\left(v_{i}^{k}(t+1)\right) \\ x_{i}^{k}(t) \text { rand } \geq T\left(v_{i}^{k}(t+1)\right)\end{array}\right.$    (13)

The simulation process of PSO stops, if converged with the predetermined stopping condition, such as highest number of iterations/best fitness value, in which the highest number defined for iteration is 100. On the other hand, if particles have surrounded by local optimum, they will reset to 0.

Defined the initial value of w as 1.4 and it changes according to the iteration, as expressed by the subsequent equation,

W = (w – 0.4) * (t_max -t) / (t_max + 0.4)    (14)

here, the highest number of iterations denoted by t_max, and the current iteration has notated by t.

3.4 Detecting epilepsy using thirteen layer convolutional neural network

The CNN comprises the essential layers, namely the convolution layer, max-pooling layer, and fully connected (dense) layer. Although the network generally inclines for efficient learning, as regards to attaining deeper, yet there is a possibility to impact on the duration of computation. To overcome this issue, we deliberated for developing the network structure that solely needs lesser time for computing. The superior performance has achieved in the process of classification that accompanies the well-regulated parameters in the event of the training phase. The convolutional layer has entwined to the input signal through a kernel (window). Post-convolution, a feature map has produced for the succeeding layer. Subsequently, the batch normalization layer has assigned to stream within the in-between layers, concerning the normalization of input data that facilitates the quicker learning and boosting. Besides, the involvement of rectified linear unit aims at the threshold of input data and diminishes the redundancies among data. Afterwards, the max-pooling layer takes place to decrease the feature map size. In the fully connected layer, each neuron of the max-pooling layer has correlated with one and all, in which the output forecasts the result (Normal/PD) of the input signal.

The implementation of the structural design of a thirteen-layer CNN surpasses the requirement of the conventional feature representation stages. Figure 2 outlines the structure of the proposed approach.

The conventional convolutional neural network was designed with thirteen layers, to match the performance metrics of Genetic based CNN. This process also includes the algorithm called Adam optimization. The activation function applied for all the layers is a piecewise linear function called Rectified Linear Unit (ReLU) function. This function overcomes vanishing gradient problem, allowing the models to learn faster and perform better. The final layer consists of Softmax function, which gives a probability of the label candidates, known as list of classes.

Based on the rendered dataset, the overall parameters are correspondingly regulated, which proves the maximum accuracy of training. The implementation of the structural design of a thirteen-layer CNN surpasses the requirement of the conventional feature representation stages.

2.png

Figure 2. Proposed thirteen layer CNN

The model has processed through two phases in order to authenticate the model, i.e. i) data training, and ii) model testing. i) The overall data has segmented as ten equal portions by presenting the stratified tenfold cross-validation, from which nine have utilized for training, and the remaining one reserved for testing the model. The iteration for ten times ensures the participation of overall portions during the stages of training and testing. ii) 20% of the cross-validation training data have designated to authenticate the model as regards the evaluation of training improvements at post-completion of each iteration. The process also includes the algorithm called Adam optimization that accompanies the learning rate of 0.0001, and applied some activation functions, especially the function named ReLU has been applied with entire layers and another function Softmax with the final layer. Subsequently, the dropout layer has been assigned with the dropout value of 0.5. Based on the rendered dataset, the overall parameters have correspondingly regulated, which proves the maximum accuracy of training. Furthermore, the Brute Force approach helps attaining the kernel size and the number of filters.