Modified Gorilla Troops Optimization with Deep Learning Based Epileptic Seizure Prediction Model on EEG Signals

Epilepsy is a common neurological condition categorized by ES and ranks second among common neurological disorders after stroke, as per a report by the World Health Organization (WHO) [1]. Seizures can occur irrespective of the host attributes or circumstances. Patients affected by epilepsy suffer from unforeseen and sudden seizures; at the time, they cannot safeguard themselves and are susceptible to injury, suffocation, or death due to traffic accidents and fainting [2]. However, this disease can be treated with surgery and medications, no cure exists, and treatments with anticonvulsants are ineffective for different epilepsy types [3]. EEG plays a significant role in epilepsy identification because it measures variances in voltage changes among electrodes on the subject's scalp by sensing ionic current flows within brain neurons and offers spatial and temporal data regarding the brain [4]. Detection using EEG requires direct investigation by a doctor, along with a significant amount of effort and time [5]. Additionally, specialists with different levels of diagnostic experience occasionally have contradictory views on diagnostic outcomes [6]. Thus, developing an automated computer-aided technique for detecting epilepsy is urgently required. Several seizure-predictive studies divide successive epileptic EEG signals into three states: ictal (period of seizure), interictal (interval among seizures), and preictal (before seizure onset). Generally, ES prediction refers to a binary classifier that differentiates between interictal and preictal [7]. Over the decade’s, numerous advanced techniques have been modeled in ES estimation. Conventionally, EEG-related SP techniques focus on classification and feature extraction [8]. The authors manually framed the discriminatory features of EEG signals, like time-frequency domain, and time–frequency domain features [9]. DL, which derives discriminatory features automatically, is widely employed. In addition, the utility of neural networks (NNs) and DL methods in expanding computer methods for disease recognition, disease detection, and many other fields has attracted the attention of many researchers [10].

The adverse effect upon quality of life caused by chronic seizures is substantial. Epilepsy accounts for more than 0.5% of the global illness burden, based on the number of years lost by persons owing to early death and years spent in suboptimal health. The costs associated with treating epilepsy, avoiding premature mortality, and losing time at work add to a sizable sum over time. According to a cost-benefit analysis conducted in India, funding for primary and secondary treatments and other medical expenses can alleviate some of the financial strain caused by epilepsy. Managing epilepsy may be more difficult than dealing with the seizures themselves because of the widespread discrimination and stigma. Individuals with epilepsy may be targets of discrimination. As a result of the negative implications of the disease, some people may be reluctant to seek help. It is anticipated that by 2050, the population of persons aged 65 years and over would have increased from 461 million to 2 billion. There will be significant consequences in the fields of social and medical healthcare because of this massive increase. Home Adaptive Response (HAR) is a powerful tool for tracking the physiological, psychosocial, and mental status of the elderly in their environments [11, 12].

During the first autonomous diagnosis, the systems achieved 76–90% accuracy, with a false detection rate of 1%–0.7% per hour. Therefore, developing methods with low computational costs is crucial. Because there is a large amount of information in an individual's EEG recording, it is vital to develop tools to oversee classification and feature selection procedures. Approaches to predicting seizures have certain limitations, including those described below.

• The key drawback of the LSTM unit of generative adversarial networks is that it might lead to right or left amplitude preponderance in EEG measurements.
• Secondly, the use of various algorithms, like the support vector machine and K-nearest neighbours, to forecast ES is hampered by a lack of directional and phase information.
• Third, the system for SP based on the EEG spike rate approach does not use DL. Consequently, this approach cannot be used to evaluate epileptic episodes.
• Fourth, the main problem with the generative adversarial network approach to SP is that it does not increase anticipation time.
• DL approaches for SP are limited by a dwindling signal-to-noise ratio (SNR) and an increase in the number of input parameters.

The problem of declining classification accuracy in SP systems has been the subject of several research studies. Because patients' EEG recordings will show seizures in different places and at different intensities, most SP algorithms must be tailored to each user. Traditional SP methods include signal pre-processing, feature selection, and classification [13-16].

1.1 Motivation

The development of DL over the last several years has increased the use of DL to diagnose epilepsy. When it comes to feature engineering, DL algorithms do not require a time-consuming human feature design procedure and have the potential to provide better outcomes than traditional methods. On the other hand, automated EEG data analysis can improve patient care by reducing the time that diagnostics take and the amount of human error that occurs. The focus of this research is on one of the first steps involved in identifying whether a person's brain activity is typical or abnormal, which is the starting point for EEG interpretation. Analyzing the signals produced by the frontal lobes is a potential method for diagnosing seizure disorders.

The driving force behind this work is to enhance the precision and efficiency of forecasting seizures using EEG signals. The present method is arduous and requires specialized knowledge, but an accurate and early diagnosis is critical for proper treatment. The MGTODL-ESP model aims to tackle these obstacles by incorporating a changed optimization algorithm and deep learning techniques. The proposed work demonstrate that the MGTODL-ESP approach yields the highest accuracy with a rate of 98.50%. Thus, this work aims to improve the existing techniques for predicting seizures and develop a model that is more accurate and efficient, thereby improving the standard of living for those who suffer from epilepsy.

Researchers are now facing several obstacles while attempting to identify seizures experienced by epileptic patients using brain imaging techniques and DL approaches.

• First, it’s important to note that just a small fraction of EEG recordings is included in the existing databases. Therefore, these signals cannot be implemented in practical settings. Therefore, it is difficult to recognize real-time warnings, and the clinical information may be inaccessible to the general public.
• Second, these enormous quantities of information cannot be combined because they have different sampling rates and frequencies. It is also impossible to train a model with such a small amount of usable data. By Diagnosis of epilepsy, patients will get specific alert messages that will be sent to physicians, relatives, hospitals, and family members using wearable gadgets, ensuring patients receive the appropriate therapy promptly.

1.2 Research objectives

This research study presented a model of EEG-based approach for forecasting ES, named “Modified Gorilla Troops Optimization with Deep Learning," was discussed in this work (MGTODL-ESP). The primary objective of the MGTODL-ESP method is to categorise across various kinds of seizures. The initial step in preprocessing the EEG data was a Z-score normalisation. The optimal characteristics were chosen using the MGTO technique. In order to forecast seizures, the grey wolf optimizer (GWO) was combined with the gated recurrent unit (GRU) framework. The GRU hyperparameters may be optimised with the support of the GWO approach.

The remaining sections of this work are arranged as follows: The current techniques for predicting seizures, as well as the drawbacks of such approaches, are discussed in Section 2. Section 3 provides a thorough explanation of the methodology proposed for the SP model. Section 4 contains the proposed SP model's results, discussion, and performance evaluation. The research is concluded in Section 5.

In this proposed work an effective feature selection model using a DL-based ES prediction method called the MGTODL-ESP model was created to use EEG signals to predict different types of seizures.  As shown in the MGTODL-ESP technique, there are four main steps: data preprocessing, MGTO-based selective feature selection, GRU classification, and GWO-based hyperparameter optimization. A block diagram of the MGTODL-ESP system is shown in Figure 1.

image003.png

Figure 1. Block diagram of proposed MGTODL-ESP approach

3.1 Data pre-processing

Initially, the presented OAOFS-DBNECD methodology converts the EEG signals into .csv format to make it appropriate to more processing. Then, the pre-processing of the EEG signals was carried out with the use of Z-score normalization [23]. It can be defined as the process of acquiring ranges of input data or normalized values from unstructured data through concepts such as standard deviation and mean. This can be established by dividing presented data of every data instance by standard deviation of all gates, as in equation, and subtracted mean of all data instances from that value. Equation was employed for mapping the values of transformed input among the target range [x, y].

$A_{Z a(t)}=\frac{A(t)-A_j(t)}{A_j(t)}$               (1)

$A_{z a d(t)}=c+\frac{(d-c)\left(A_{z s}\quad(t)-A_{z s \min\quad }\right)}{A_{z s \max }\quad-\quad A_{z s \min }}$              (2)

The scaling values of z‐score normalized field $A_{Z a(t)}$ were shown here as $A_{Z a d(\mathrm{t})} \cdot A_i(t)$ and $A_j(t)$, which are computed from the trained data sets, denoted the standard deviation and mean at each gate time $t$, where $A_{Z \max }$ and $A_{Z S_{\min }}$ were absolute final and initial gate values noted for all provided gate periods in test datasets.

3.2 Feature selection

Feature selection refers to the process of narrowing down a large pool of potential features (predictive variables) to a manageable number. As a result of this process, the model should be more accurate and interpretable, with less overfitting and more efficient computing. Several methods exist to accomplish this goal, including dimensionality reduction, feature importance calculation, and correlation analysis.

Here once the preprocessing stage is completed, then implementing a method called feature section. In this stage, reduction of features takes place and it also makes an attempt to forecast which characteristics are more significant and relevant to our proposed model.

3.2.1 Steps involved in MGTO based feature selection

The MGTO (Modified Gorilla Troops Optimization) algorithm was used to select the best features. Gorilla troop optimization (GTO) is a technique based on swarms. It was inspired by the social life of the gorilla, which was the biggest primate on Earth [24]. There are two parts of this method: exploitation and exploration. Five different operators attempt to copy the optimization function for a gorilla's behavior. During the exploration phase, there were 3 operators: moving to a known position, moving to a new position, and moving to other gorillas. During the exploitation phase, the search was run by two operators who tracked the silverbacks and competed for adult females.

Silverback is in charge of a group of gorillas and he can do everything. Gorillas sometimes go to places in nature that are new to them, or that they have been to before. In each step of the optimization process, the best candidate solution was called a "silverback solution.” Eq. (3) is used to represent the three strategies in the exploration phase, where $p$  refers to the variable within [0,1] and is applied to choose the migration strategy for an unknown position. if rand<p, the algorithm for migration to an unknown location is selected. While p is a variable in the range [0,1] used to determine the best migration technique for a location whose exact coordinates are uncertain. if rand<p, the algorithm for migration to an unknown location would be selected. Next, when rand≥0.5, the next model, that of movement towards remaining gorillas, would be chosen. The algorithm for migration to a known location is preferred when rand<0.5,

$\begin{aligned} & G X(t+1)  = \begin{cases}(U B-L B) \times r_1+L B & \text { rand }<p \\ \left(r_2-C\right) \times X_r(t)+L \times H & \text { rand } \geq 0.5 \\ X(i)-L \times\left(L \times\left(X(t)-G X_r(t)\right)+r_3 \times(X(t)-G X(t))\right) & \text { rand }<0.5\end{cases} \end{aligned}$                            (3)

whereas GX(t+1) represents the candidate location vector of gorillas in the second iteration, and X(t) indicates the existing location vector of the gorilla. Furthermore, r₁, r₂, r₃ and rand specifies a randomized value between 0 and 1. UB and LB represents the upper and lower boundaries, respectively. X_r and GX_r denote the candidate location vector of gorillas chosen randomly

$C=F \times\left(1-\frac{I t}{M a x l t}\right)$                    (4)

In Eq. (4), $I t$ and MaχIt represent the existing and maximal iteration values, respectively. Initially, the variation value is produced in a larger interval, and later the rehabilitated interval of the variations value might be reduced at the end of the optimization phase:

$F=\cos \left(2 \times r_4\right)+1$                    (5)

In Eq. (5), r₄ denotes the random number within [-1,1]:

$L=C \times l$                         (6)

Eq. (6), l indicates a random number within [0,1]. Furthermore, Eq. (6) was applied to simulate the silverback leadership. Due to a lack of experience, silverback gorilla has difficulty in making the right decision to manage the group or find food. However, they can attain extreme stability and adequate experience with the leadership method. Furthermore, H in Eq. (3) can be evaluated using Eq. (7). Z in Eq. (7) can be evaluated using Eq. (8), whereas Z denotes a random number within [-C,C].

$H=Z \times X(t)$                        (7)

$Z=[-C, C]$                       (8)

In the last exploration, a group function was implemented to calculate the cost of each GX solution. Once the cost is recognized as (t)<X(t), the X(t) solution is substituted with the GX(t) solution. Thus, a better solution can be considered as silverback.

In the exploitation phase, competition for adult females and the two behaviors of the silverback were adopted. If the silverbacks and other gorillas were younger, they implemented their duties well. For example, a male gorilla follows a silverback. Furthermore, every member influence other member. When C≥W, the following approach can be implemented.

$G X(t+1)=L \times M \times\left(X(t)-X_{\text {silverback}}\right) X(t)$                       (9)

In Eq. (9), X_silverback indicates the vector of silverback that present the optimum solution:

$M=\left(\left|\frac{1}{N} \sum_{i=1}^N G X_i(t)\right|^g\right)^{\frac{1}{g}}$                       (10)

In Eq. (10), GX_i(t) denotes the vector position of the candidate gorilla at time t; N denotes the total number of gorillas;

$g=2^L$                       (11)

Adolescent gorillas have a significant phase of puberty during which they compete with other males for the attention of females. This kind of rivalry is often intense, may last for many days, and has an impact on those around it.

$\begin{aligned} G X(i)=X_{\text {silverback }} -\left(X_{\text {silverback }} \times Q-X(t) \times Q\right) \times A\end{aligned}$                       (12)

$Q=2 \times r_5-1$                       (13)

$A=\beta \times E$                       (14)

$E=\left\{\begin{array}{l}N_1, \text { rand } \geq 0.5 \\ N_2, \text { rand }<0.5\end{array}\right.$                       (15)

Eq. (15), Q was adopted to simulate the effect, which was evaluated using Eq. (13). Furthermore, r₅ indicates a random number between 0 and 1. Eq. (14) an be utilized for calculating the coefficient vector of the degree of violence in conflict, whereas β indicates the variable that should be provided beforehand in the optimization technique. E can be used to stimulate the effects of violence in the solution dimension. When rand≥0.5, E is equivalent to random values in the uniform distribution and problem dimensions.

To accelerate the converging speed of the GTO algorithm and fortify the optimization capability, the MGTO technique was modeled using the opposition-based learning (OBL) method for screening N initial populations. The steps of the multistage population initialization algorithm joining OBL and logistic mapping are as follows:

1) leverage logistic mapping for generating N the number of individuals to form original populationN1;

2) Use Eq. (16) for finding an opposed solution for all individuals to form a reverse population N2;

$\tilde{X}_i^d=U_b^d-L_b^d+X_i^d$                       (16)

Equation X_i^d signifies the reverse individual of initial individual. Eventually, combine the reverse population, and original population, compute the fitness values and chooses the first N amount of individuals with small fitness to form final initial population.

The fitness function (FF) of the MGTO considered the classifier accuracy and count of selected features. It optimizes the classifier accuracy and decreases the selected features set size. Hence, the following FF was employed for assessing individual solutions, as given in Eq. (17).

Fitness $=\alpha *$ ErrorRate $+(1-\alpha) \frac{\# S F}{\# A l l\_ F}$                     (17)

whereas Error Rate means classifier error rate exploiting the selected features.

3.3 Seizure prediction process

In this work, the GRU model is applied to predict epileptic seizures. RNN can be a kind of ANN that was suitable to investigate and process time sequence datasets, in contradiction to traditional NN [25]. It depends primarily on the weight connections amongst the layers. The RNN applied the hidden layer (HL) to maintain dataset in previous instant, and output can be affected by current conditions and prior memory. Here, $x^{(t)}$ and $\hat{y}^{\langle t)}$ represents input and output at time t, $a^{(t)}$ signifies the outcome of single HL at time t, and $\omega_{a a}^{\langle t\rangle}$, $\omega_{a x}^{\langle t\rangle}$ , and $\omega_{a y}^{\langle t\rangle}$ correspondingly shows HL, input, and output weighted matrices.

$\begin{gathered}a^{\langle t\rangle}=g_1\left(\omega_{a a} a^{\langle t-1\rangle}+\omega_{a x} x^{\langle t-1\rangle}+b_a\right), \\ \hat{y}^{\langle t)}=g_2\left(\omega_{a x} a^{\langle t\rangle}+b_y\right),\end{gathered}$                     (18)

whereas b_a and b_y correspondingly shows the bias vector of single HL and output. g₁ and g₂ signifies nonlinear activation function. The RNN accomplishes well when the outcome was neighboring to relevant input; although, when the time interval was lengthy and weight count becomes large, input would be smaller effect on output due to gradient vanishing problems. The GRU framework has been demonstrated in Figure 2, whereas $\sigma$ and tanh signify activation function, $c^{\langle t-1\rangle}$ represent input of existing unit, that is outcome of previous unit, $\hat{y}$ specifies output of existing unit, which interconnects to input of subsequent unit. $x^{(t)}$ signifies input of trained dataset, $x^{(t)}$ specifies results of unit, generated by activation function, $\Gamma_r$ and $\Gamma_u$ correspondingly represents reset and upgraded gate, and candidate activation $\tilde{\boldsymbol{C}}^{\langle t\rangle}$ can be equally evaluated. It comprises of 2 gates in GRU, i) upgrade gate, which continues previous dataset to the current state; the values of $\Gamma_u$ ranges from 0 to 1, the nearby $\Gamma_u$ was to 0, the other previous dataset it conserves; another was reset gate that can be applied to describe whether the existing state, as well as previous dataset, are incorporated or not. The value of $\Gamma_r$ range in -1 to 1, while value gets lesser of $\Gamma_r$ the other previous dataset get ignores:

$\begin{gathered}\Gamma_u=\sigma\left(\omega_u\left[c^{\langle t-1\rangle}, x^{\langle t\rangle}\right]+b_u\right), \\ \Gamma_r=\sigma\left(\omega_r\left[c^{\langle t-1\rangle}, x^{\langle t\rangle\rangle}\right]+b_r\right), \\ \tilde{c}^{\langle t\rangle}=\tanh \left(\omega_c\left[\Gamma_r * c^{\langle t-1\rangle}, x^{\langle t\rangle}\right]+b_c\right), \\ c^{\langle t\rangle}=\left(1-\Gamma_\nu\right) * c^{\langle t-1\rangle}+\Gamma_\nu * \tilde{c}^{\langle t\rangle},\end{gathered}$                  (19)

whereas, ω_u, ω_r, and ω_c correspondingly indicates trained weighted matrices of upgrade and reset gate and candidate activation $\tilde{\boldsymbol{C}}^{\langle t\rangle}$, and b_u, b_r, and b_c denotes bias vector.

image044.jpg

Figure 2. Framework of GRU

To modify the hyperparameter values of the GRU method, the GWO is used. Mirjalili et al. developed a new SI optimized technique named GWO [26]. Indeed, it is original method that accelerates hunting and social hierarchies of GW. To develop the social performance of GW, it is classified into 4 states namely α, β, $\delta$, and $\omega$. α considering that optimal solution implemented by $\beta$ and $\delta$, correspondingly, and residual solution derives in $\omega$. The first 3 fittest wolves called α, $\beta$, and $\delta$ are nearby the prey support ω to distinguish the food from complex areas. In the surrounding stage, wolf increases the place of, $\beta$, or $\delta$ as demonstrated in Eqns. (20) and (21):

$\vec{D}=\left|\vec{C} \cdot \vec{X}_{p(t)}-\vec{X}(\mathrm{t})\right|$                           (20)

$\vec{X}(\mathrm{t}+1)=\vec{X}_{p(t)}-\vec{A} \cdot \vec{D}$                           (21)

Now t shows the existing iteration, $\vec{X}_{p(t)}$ signifies the present location of prey and $\vec{X}(\mathrm{t})$ implies the present place of wolves. $\vec{D}$ denotes the distance amongst wolf and prey, and co-efficient vectors $\vec{A}$ and $\vec{C}$ resulting in mathematical processes in the following.

$\vec{A}=2 \vec{a} \overrightarrow{r_1}-\vec{a}$                         (22)

$\overrightarrow{\mathrm{C}}=2 \overrightarrow{r_2}$                         (23)

where, $\overrightarrow{r_1}$ and $\overrightarrow{r_2}$ implies the 2 vectors lie within [0,1], the component of $\vec{a}$ has linearly reduced from 2 to 0. Now, α, β, and δ describes location nearby to prey locations. In the process of hunting, topmost 3 are optimal solutions and residual wolves ω are suitable for substituting on the first 3 optimal wolves:

$\overrightarrow{\mathrm{D}}_\alpha=\left|\overrightarrow{\mathrm{C}}_1 \cdot \vec{X}-\vec{X}\right|$                         (24)

$\vec{D}_\beta=\left|\vec{C}_2 \cdot \vec{X}-\vec{X}\right|$                         (25)

$\vec{D}_\delta=\left|\vec{C}_3 \cdot \vec{X}-\vec{X}\right|$                         (26)

$\overrightarrow{\mathrm{X}}_1=\vec{X}_a-\overrightarrow{\mathrm{A}}_2 \cdot\left(\overrightarrow{\mathrm{D}}_\alpha\right)$                         (27)

$\vec{X}_2=\vec{X}_\beta-\vec{A}_2 \cdot\left(\vec{D}_\beta\right)$                         (28)

$\vec{X}_3=\vec{X}-\vec{A} \cdot\left(\vec{D}_\delta\right)$                         (29)

$\vec{X}(\mathrm{t}+1)=\frac{\overrightarrow{X_1}+\vec{X}+\overrightarrow{X_3}}{3}$                         (30)

whereas $\vec{X}_\alpha$ indicates the location of $\alpha$, $\overrightarrow{\mathrm{X}}_\beta$ determined the location of $\beta, \overrightarrow{X_\delta}$ denotes the location of $\delta, \vec{X}$ shows the location of current solutions, $\overrightarrow{\mathrm{C}_1}, \overrightarrow{\mathrm{C}_2}$ and $\overrightarrow{\mathrm{C}_3}$ suggests the vector produced arbitrarily. At this time, $\overrightarrow{\mathrm{A}_1}, \overrightarrow{\mathrm{A}_2}$ and $\overrightarrow{\mathrm{A}_3}$ indicates arbitrary vector, and $t$ indicates the iteration count. The step size of ω wolf implemented then α, β, and δ are demonstrated in Eqns. (24)-(26). Later, the resulting location of ω wolf has assessed by Eqns. (27)-(30).

The GWO method makes a derivate of fitness function (FF) to have greater classifier outcome. It defines a positive value for designating higher outcomes of candidate solutions.

In this article, the minimal classifier rate of errors indicates the FF, as given in Eq. (31).

$\begin{aligned} & \text { fitness }\left(x_i\right)=\text { ClassifierErrorRate }\left(x_i\right) = & \frac{\text { number of misclassified samples }}{\text { Total number of samples }} * 100\end{aligned}$                         (31)