Improved Scalp Swarm Optimization with Generative Adversarial Network for Recognition of Motor Imagery EEG

Abstract

Motor imaging (MI) has been commonly employed in the domains of nervous analysis and robot control as an essential model of impulsive brain-computer interfaces (BCIs). Several approaches for extraction and classification based on MI signals have presented by researchers. Because of its random initialization technique, deep-learning (DL) procedures like convolutional neural networks (CNNs) employed in motor imagery categorization would suffer from the problem of extracting the features and improving the classification performance. To overcome these shortcomings, the proposed work presented a technique for reconstructing MI signals using empirical mode decomposition (EMD). In which it can manage the non-stationary problem and mix their Intrinsic Mode Functions (IMFs) extended to multichannel analysis. The proposed works uses the transformation technique of discrete wavelet transform (DWT) for extracting the signal. In classification wise, the proposed work used conditional generative adversarial network (CGAN), which provides the better classification performance and reduces the computational time. The proposed work used both binary class and multi-class classification. Scalp swarm optimization (SSO) used to enhance and optimize the learning parameters of the GAN model because in multi-class subjects, the performance of CGAN gradually degrades. In this BCI experimentation, the proposed work used two BCI competition datasets, such as BCI competition three dataset III (a) and BCI competition three dataset IV (a). Furthermore, evaluate the proposed technique performance by evaluating and comparing with DL technique as CNN (Alexnet and Resnet) model. The CNN achieved a classification accuracy performance of 81.61% in multiclass, while the CGAN showed better performance at 89.11%.

Keywords:
Brain computer interfaces (BCI); Motor imaging; Deep learning; Conditional generative adversarial network (CGAN); Slap swarm optimization (SSO)

HIGHLIGHTS

Generative adversarial network provides the better classification performance.

CGAN showed better performance at 89.11% and reduces computational time.

SSO is used to enhance and optimize the learning parameters of the GAN model.

INTRODUCTION

A brain-computer interface (BCI) is a communiqué scheme that ensures to the brain's typical output channels of peripheral muscles and nerves. The general architecture of BCI is as shown in Figure 1, which is divided into four main components which comprise signal acquisition, preprocessing, feature extraction, classification to finally the translation of commands, and feedback. In this, the signal acquisition is important for recording the brain waves and sending them to preprocessing for signal enhancement and further to reduce the noise [¹, ²]. Feature extraction generates the discriminative signal to improve the original signal. The classification is used in addition to lessening the size of the data applied and to provide features for the device command. Finally, the feedback signal, which will be in the form of audio or visual, helps the subject to increase the accuracy. In general, an electroencephalogram (EEG) is a visual representation of brain electrical potentials recorded onto a paper during a specified period. The EEG signal was captured using an electrode and a computer storage device. The electrodes and leads are inserted into a head box or jack box, which is linked to the EEG machine through a screened multi way input connection. Sockets corresponding to each electrode site are typically identified and placed in accordance with the 10-20 electrode situation schemes. The electrode position must be established by measuring from a standard landmark on the scalp [¹]. The measurement should be proportionate to the size and form of the skull. The nature of the EEG signal is analogy and non-stationary. The analog signal is converted into a digital signal and stored in computer memory, and the stored signal proceeds with the processing with classification [⁴]. The processing of raw signals for classification suffered from classification ratio and selection of feature components. Extraction of features and selection of features play an essential role in the classification of motor imagery electroencephalography (MI-EEG) data sortin [⁵]. The practice of feature extraction applied various frequency-based transforms with different levels of decomposition and signal range analysis.

Empirical mode decomposition (EMD) is a relatively new approach for handling non-linear and stationary signals [⁶]. Additional data analysis approaches, such as Fourier and wavelet-based approaches need the use of predefined basis functions in order to describe a signal. The EMD is a complete signal-based process that does not require any prior knowledge. The EMD's objective is to represent a signal as a set of adaptively determined basis functions with well- definite frequency localization planes. The chief impression behind this practice is to partially rebuild the signal, with the relevant Intrinsic Mode Functions (IMFs) matching to the signal's most essential structures (low-frequency components). The filtering approach is based on the evidence that the mainstream of the signal's significant structures frequently focus on the lower frequency ones (latest IMFs) and diminish towards the high-frequency modes (first IMFs)[⁷]. Most classification methods compromise on the feature parameter and channel range choices. The kernel technique and threshold function of an artificial neural network were used to choose the features and parameters. While the kernel mechanism provides the architecture of classification algorithms, the threshold function determines the selection range of features. The EEG data processing process makes use of the majority of neural network prototypes. Rapid and precise sampling of EEG data is made possible by neural network models' effectiveness in processing and classifying EEG data [⁸]. The majority of the current model's concentration is on static data, hence it is unable to accurately classify quickly changing brain signals. For instance, the present range of motor imaging EEG categorization accuracy is between 60% and 80%. Innovative learning techniques are needed in BCI systems to handle dynamic data streams [⁹]. Deep learning (DL), which has recently been widely used in BCI applications, has shown effective in tackling the aforementioned issues. DL has two advantages. To begin with, it doesn't require any time-consuming preprocessing or feature engineering because it functions directly on raw brain impulses. Second, deep neural networks can contain both representative high-level traits and hidden associations by exploiting deep structures. DL algorithms that adaptively train discriminative features to classify input data into predefined categories. By classification based on probabilistic prediction and nonlinear change, discriminative algorithms identify distinctive properties. Recurrent neural networks (RNN), CNN, and other discriminative algorithms can be utilized for both feature engineering and classification [¹⁰]. In BCI, generative algorithms are frequently used to reconstruct or generate a collection of brain signal samples for expanding the training set. In this field, Generative adversarial networks (GANs) are widely used models. Additionally, the accuracy of the classifier model can be improved by optimizing the model using various optimization techniques, such as particle swarm optimization (PSO), ant colony optimization (ACO), and other approaches.

Scalp Swarm Optimization (SSO): Scalp Swarm Optimization is likely a specific optimization algorithm or technique. Optimization techniques are used in machine learning and other fields to find the best possible solution to a problem. SSO could be a variant of particle swarm optimization (PSO) or another optimization algorithm.

Generative Adversarial Network (GAN): A Generative Adversarial Network is a type of neural network architecture that consists of two networks, a generator and a discriminator, which are trained together in a competitive manner. GANs are often used in generative tasks, like generating images or data that resemble a certain distribution.

Recognition of Motor Imagery EEG: This part of the title suggests that the research or project involves processing EEG signals related to motor imagery. EEG is a method for recording electrical activity in the brain, and motor imagery is the mental representation of a specific movement without actual physical movement. Recognizing motor imagery from EEG data can have applications in brain-computer interfaces and neuroscience.

Improved: This indicates that the research aims to enhance or improve existing methods or approaches related to the topics mentioned in the title. It suggests that the study may introduce modifications or optimizations to existing techniques for better performance.

The study discussed some existing work on motor imaginary task recognition, in the past decades, there are several researchers are focusing motor imaginary task recognition by using different models, which are followed as; the GAN for EEG Signal have been a few researches on the usage of GANs for EEG [¹¹]. It is challenging to obtain superior EEG data due to the stringent criteria for individuals and experimental settings. GANs were trained using EEG characteristics to enhance picture production. GANs have led to significant advances in the generation of time-series data in fields such as voice processing. Given the restricted data collection and insufficient concentration of participants during testing, it is critical to gather sufficient training data and relevant characteristics for a prospective end-user of a BCI system. To address this problem, the author proposed combining a conditional vibrational auto-encoder network (CVAE) with a GAN for learning latent illustrations from EEG brain signals [¹²]. The benefits of both statistics and feature matching in making the training procedure faster and more reliable, as well as addressing the difficult of small-scale datasets in DL for motor imaging tasks, are discussed. This capability of GANs can be particularly beneficial for BCIs because gathering a large number of samples might be costly and time-consuming. As a result, the author employed the GAN model, which can capture significant aspects of motor imagery. Separate GANs are trained to produce fake EEG samples that correspond to the two kinds of trials in the data set.

Extraction of EEG will be many ways for extracting the waveforms of the EEG signal. γ (0.5-4 Hz), θ (4-8 Hz), α (8-12 Hz), β (12-30 Hz), and δ (above 30 Hz) are the frequency bands that the brain's activity is split into [¹³]. The authors used discrete wavelet transform (DWT) and EMD to break down the EEG signal into a collection of stationary time series IMFs by first decomposing it into a sum of narrow band signals and then the subband signal. Following that, the suitable IMFs for signal reconstruction are chosen. However, it is a difficult problem to choose the best IMFs of EMD for extracting the characteristics of EEG signals, but must choose the best IMF of EMD for diagnosing sleep disorders using EEG brain signals [¹⁴]. CGAN for MI-EEG Signal n research, a novel version of GANs was reported. The first improvement to the original GANs was CGANs, which employed deep CNN for both the generator and the discriminator for improved training [¹⁵]. The deep neural network-based decoding model has received a lot of interest in the field of MI signal processing. It is challenging to obtain large-scale and high-quality EEG data due to the rigorous criteria for participants and experimental settings. However, the enactment of a DL is directly proportional to the amount of the dataset. The authors demonstrated that the CGAN produced high-quality fake EEG spectrogram data [¹⁶].

Regular convolutional neural networks were purposefully chosen because, on the one hand, the majority of GAN [¹⁷] studies use the CGAN architecture, which is based on CNNs, [¹⁸] and on the other hand, the local and hierarchical structure of CNNs may allow for better interpretability, which is important for brain signals in a neuroscientific or clinical context. In classification and recognition of MI-EEG signal is challenging for every researcher, due to low classification accuracy. So different authors have used different optimization techniques to improve the accuracy. To optimize the classifier parameters and the author have to extract the most significant characteristics [¹⁹]. A hybridization of particle swarm optimization (PSO)-based rough set feature selection is presented. However, in this proposed study, Scalp swarm optimization is used to improve the classifier performance and conduct the qualitative analysis by using different optimization techniques, which are mentioned in the result unit.

The most significant issue challenge is properly predicting the EEG signals and extracting the hidden features from the signal. A neurological disorder can develop in the human brain as a result of aberrant EEG flow. The first step is to extract and analyze features from a non-stationary signal, such as an EEG signal. Another, the proper classification of EEG signals. However, most of these studies, they implemented two-class classification only, not a multi-class classification. Some studies have used multi-class classification, but the performance was not better due to the learning rate and improper selection of hyper parameter of the classifier model. Solutions of the above problem are as follows, first of all, establish a well-defined and well-structured process for extracting the hidden features from a very transient and non-stationary signal like EEG signal by using EMD technique. For this, signal transformation method called as DWTwas used [²⁰]. Additionally, we established an efficient classifier model that can recognize and classify the EEG signal. Here we implement the multiclass classification by using CGAN with additionally Scalp swarm optimization technique used to optimize the classifier to fine-tune the accuracy performance.

In recent times, deep learning has been extensively used in BCI applications and has demonstrated effectiveness in resolving the aforementioned problems. Discriminative algorithms, such as recurrent neural networks (RNN), CNN, and others, can be used for both feature engineering and classification. Generative algorithms are widely employed in BCI to reconstruct or create a batch of brain signal samples to advance the training set. Generative adversarial networks (GANs) are popular models in this area. And improving the classifier model accuracy by optimal model by using different optimization are used in existing PSO, ACO, and different methods.

The primary contribution of this paper, similar to convolutional neural networks (CNNs) employed in motor imagery categorization, is to address the challenge of feature extraction and improve classification performance. To overcome these shortcomings, the proposed work presents a technique for reconstructing MI (motor imagery) signals using empirical mode decomposition (EMD), which effectively handles non-stationary problems and combines their Intrinsic Mode Functions (IMFs) for multichannel analysis. The proposed approach employs the discrete wavelet transform (DWT) for signal extraction

METHODS

The proposed approach architecture and an overall outline are deliberated in this section, which is labelled in Figure 2. The online experiments demonstrate that the planned intelligent wheelchair structure can precisely judge the instructions of the subjects. It proves that the intelligent wheelchair scheme created is a viable BCI application. In this architecture describes, initially an EEG is preprocessing by removing and reducing the background noise from EEG signal. Then the preprocessed signal is given to the EMD, analysis of nonlinear and non-stationary signals is also used for feature extraction. EMD is realistic to the signals, and it separates the signal into residual elemental IMFs (C3, C4, and Cz), which offer instantaneous frequency as functions of time, allowing for good recognition of embedded structures. Decomposed IMF’s are combined to average and it is applied to the DWT technique, which extracts the signals in four sub bands as low-high, low-low, high-high and high low. After consider to average these four sub-Band signals in the classification process. In the classification process, we used generative adversarial network (GAN) model, in the beginning of classification, prediction accuracy is low due to less learning parameter. By increasing the learning rate and get the better prediction accuracy by using Scalp swarm optimization technique. In this experimental process, we used two datasets as BCI competition three dataset III(a) and BCI competition three dataset IV(a), which are deliberated in the following subsections.

Figure 1
The general architecture of BCI

Figure 2
Proposed methodology flow diagram

Dataset

In this experimentation study, we have used two diverse BCI competition datasets such as BCI competition three dataset III(a) and BCI competition three dataset IV(a). Both dataset descriptions are described in the following subsections with the help of Table 1. As datasets are used, all methods were performed in accordance with the relevant guidelines and regulations. The input images used in this research work are collected from open access publicly available datasets. It is available on https://www.bbci.de/competition/iii/#datasets for experimental purposes.

Table 1
Dataset specification

Dataset	Dataset source	Channel	Subject	Class	Sampling frequency (hz)
1	BCI III(a)	4	3	4	250
2	BCI IV(a)	118	5	2	100

EEG Signal Model

The Eq. (1) represents the characteristics of the EEG signal, which represents neural activity in the brain and comprises rhythmic and evoked responses.

$X(t)=X_R(t)+X_E(t)$ (1)

Where $X_R(t)$ results from position of potentials produced by cerebral actions, $X_R\left(t\right)=M_B{S}_A\left(t\right).$The component $X_E\left(t\right)$ corresponds to evoked potentials in response to obtainable stimuli, $X_E(t)=N_B{S}_B(t)$ where $M_B$ and $N_B$ are mixing matrices and $S_A(t),S_B(t)$ are contributive elicited potentials to rhythmic and evoked responses respectively.

Electrode selection

In electrode, selection is an important process for recognition of EEG signal. There are different channels are available in the position, but in this study, we have selected three channels such as C3, Cz, and C4, which contain sufficient signals for our experiment. By selecting this electrode, after that, the proper EEG signal is taken from the electrode and it is converted into a 32*32 matrix. In this research, work directed numerous experiments on two different types of datasets. There were five subsets in the dataset. Two-class and multi-class (k3b, k6b, L1b) data from several healthy participants were used to create each of the five datasets. MI tasks included right hand movement, right foot movement, foot and tongue movement, and a combination of the two [²¹].

Feature extraction

After the electrode selection, the following step is to extract the necessary information of the processed EEG signal known as features. The set of features are generally combined into vectors called as feature vectors. This feature helps to categorize the EEG signal, thereby contributing to the performance of the system.

Empirical Mode Decomposition

Each IMF component represents a different element of the signal; for example, the first function captures the high frequency component of the original signal, and as a result, a signal representation is built from rapid oscillations to slow oscillations. The IMFs are also known as modes or components, and they are acquired by a process known as sifting. Depending on the parameters, the signal might be divided into fewer or more components. This is referred known as decomposition resolution, and it is analogous to utilizing multiple window widths in the short time Fourier transform (STFT) or selecting different wavelets in the DWT. For Gaussian noise, the EMD acts as a "wavelet-like" dyadic filter bank. Reconstructing the data from the IMF components (including the residual trend) from the longest to the shortest periods is doable. The difference between the reconstructed data from the total of all IMFs and the original data is quite tiny, indicating that the decomposition is complete.

The decomposed IMF levels will distinct as the frequency components occurring in the noisy EEG signal. If the original signal $x\left(t\right)$decomposed by the EMD can be stated as follows in Eq. (2).

$x\left(t\right)={\displaystyle \sum }_{i=1}^N{C}_i(t)$ (2)

Where the frequency of $C_1\left(t\right),C_2\left(t\right),C_3\left(t\right)\dots C_n\left(t\right)$is decreasing. So the low frequency signal levels can be written in Eq. (3).

$x_L\left(t\right)=={\displaystyle \sum }_{i=1}^N{C}_i\left(t\right),1\le L\le N$ (3)

The high frequency signal levels can be expressed in Eq. (4).

$x_H\left(t\right)={\displaystyle \sum }_{i=1}^N{C}_i(t),1\le L\le N$ (4)

When moving from one residual to the next, the number of extrema is reduced by construction. It ensures that thorough decay is accomplished in a finite number of phases, and thus the equivalent spectral supports are predictable with reduction as a result. While modes and residuals might be intuitively interpreted as “spectral,” it is important to note that their high vs. low frequency discrimination relates only locally and does not correlate with any pre-gritty sub-band filtering. Apply EMD to a noisy EEG signal to obtain various IMFs. The decomposed signal can be signified as the sum of several IMFs and a single residue in Eq. (5).

$x\left(t\right)={\displaystyle \sum }_{i=1}^N{C}_i(t)+r(t)$ (5)

where, $r\left(t\right)$ the residue signal

Notice the data set of both maxima and minima of $x(t).$

Produce the upper and lower envelopes $e_{\min }\left(t\right)and{e}_{\max }\left(t\right),$ correspondingly by involving max and min distinctly with cubic spline interpolation.

Determine the local mean as $a(t)\frac{e_{\min }\left(t\right)and{e}_{\max }(t)}{2}$

Extract the detail $h_1\left(t\right),=x(t)-a(t).$

Decide whether $h1\left(t\right)$ is an IMF or not by inspection the two basic circumstances as labelled above.

Repeat stages (a) to (e) and end when an IMF$h_1\left(t\right)$is gained.

Once the first IMF is derived $c_1\left(t\right)=h_1\left(t\right),$which is the minimum temporal scale in $x(t).$ To find the rest of the IMFs, produce the residue $r_1\left(t\right),$ of the data by subtracting $c_1\left(t\right),$as $r_1\left(t\right)=x\left(t\right)-c_1\left(t\right)$. The filtering procedure described above will be repeated until the final residue is a constant, a monotonic function, or a function from which no further IMFs can be generated. The original signal $x(t)$ is represented at the conclusion of the breakdown in Eq. (6).

$x\left(t\right)={\displaystyle \sum }_{m=1}^M{c}_m\left(t\right)+rM(t)$ (6)

Where M is represent as the sum of IMFs, $c_m\left(t\right)$ is the mth IMF, and $rM(t)$ is represent as the final residue. Each IMF in equation (1) is considered to produce a meaningful local frequency, and no two IMFs produce the same frequency at the same moment. The Eq. (7) may therefore be written as follows.

$x\left(t\right)\approx {\displaystyle \sum }_{m=1}^M{A}_m\left(t\right)+\cos [{\varnothing }_m\left(t\right)]$ (7)

Where $A_m$instantaneous amplitude and ${\varnothing }_m$instantaneous phase. Then, $C\mathrm{3,}C4andCZ$ average IMF signal is given by the Discrete Wavelet Transform, which process is signified in the following section.

The Discrete Wavelet Transform

The discrete wavelet transform was employed in the feature extraction procedure. The STFT is ineffective for assessing non-stationary data like EEG. This is due to the fact that STFT has a consistent resolution at altogether frequencies. The wavelet transform technique, which employs multiresolution, is used to examine distinct frequencies with varying resolutions. The DWT offers no redundant, incredibly effective wavelet form that may be built using a simple recursive filter approach. The number of coefficients it can produce without losing information is equal to the number of samples in the original neuroelectric waveform. The DWT enables the accurate reconstruction of the original neuroelectric waveform by employing an inverse filtering technique. The discrete wavelet transform is a transformation function that was derivate from the mother wavelet transform. The detail decompositions of transform represent with$N$. The value of $N$ defines the level of decomposition and extraction level of EEG signal data. By maintaining the temporal information included in the coefficients, the suggested feature extraction approach fully uses the DWT's time-frequency analysis. The extracted feature from the $c3IMF\mathrm{1,}czIMF1\&c4IMF\mathrm{1,}$and finally added absolute signal to the IMF signal. By utilizing a single function to decompose such signals into numerous functions. This function is known as the mother function, and it is provided in Eq. (8).

$\psi \left(t\right)=\frac{1}{\sqrt{2}}\psi \left(\frac{t-y}{x}\right)\dots .x,y\in s,c>0$ (8)

Where $xandy$ represent the scaling and shifting parameters, respectively, and $S$ represents the wavelet space The wavelet transform is depicted in Eq. (9).

$F\left(x,y\right)=\frac{1}{\sqrt{x}}\underset{s}{\overset{.}{{\displaystyle \int }}}\psi \left(\frac{t-y}{x}\right)dt$ (9)

The proposed work used DWT in this study because it gives an extremely efficient wavelet representation. Low- and high-pass filters are often used in first-level decomposition to get the representation of the digital signal as approximation components. The Eq. (10) specifies DWT decomposition,

$f\left(t\right)={\displaystyle \sum }_{k=-\infty }^{k=+\infty }{C}_{n,k}\varnothing \left(2^{-n}t-k\right)+{\displaystyle \sum }_{k=-\infty }^{k=+\infty }{\displaystyle \sum }_{k=-\infty }^{k=+\infty }{2}^{\frac{1}{2}}{d}_{j,k,}\psi \left(2^{-j}t-k\right)$ (10)

Where the approximation and detail coefficients are denoted by $d_{j,k,}$ and $C_{n,k}$ respectively, while the scale is given by and. Using a decomposition algorithm, the first approximation is broken down into smaller components. The total number of decomposed signals is n+1 at the end of the operation. db4 was chosen because it has the best quality for appropriately identifying signal features for use as the mother wavelet in this study.

CLASSIFICATION

After the feature extraction technique, the next step is to classify the features assigned to a class automatically and translate them into a command. A number of classification algorithms is used in BCI and they are termed as classifiers.

A GAN is a DL framework taught via an adversarial system. To aid in the training process, synthetic data that follows to the original distribution is created. A GAN, unlike other DL prototypes, is made up of two portions such as a generator and a discriminator. In GANs, the training is done by the generator and discriminator, and D(x) reflects the likelihood that x belongs to the class of real images rather than false examples. Simultaneously, model G attempts to create samples that are getting closer and closer to the genuine pictures.

Figure 3
Proposed Generator and discriminator GAN architecture model

The discriminator's primary objective is to calculate a probability close to 1 when the input data consists of real photographs. It is the discriminator's job to judge and learn what makes real samples different from phony ones when using fake samples. As $\text{D(G(z))}$ approaches zero, the generator attempts to go closer. It is important to train GAN model D during the learning phase and thus it can correctly distinguish between legitimate and bogus input sources. The generative network $G(\cdot )$ is commonly represented by CNN. The discriminative network $D(\cdot )$ represents the classification probability of input data. The expectation function $E(\cdot )$. For the specified training datasets, during training, the discriminator aims to enhance $V(D,G)$ whereas the generator aims to reduce $V.(D,G).$ If you've calculated the gradient, it looks like this in Eq. (11).

$g_D={\nabla }_{\theta d}\frac{1}{m}{\displaystyle \sum }_{i=1}^m\left[\log D(x|y^{(i)})\right]+\log (1-D(G(z|y^{(i)})))$ (11)

$g_D={\nabla }_{\theta d}\frac{1}{m}{\displaystyle \sum }_{i=1}^m\left[\log (1-D(G(z|y^{\left(i\right)}))\right]$ (12)

By using Eq. (12), the discriminator is trained to produce a high probability of genuine data x and a low likelihood of created data. $G(Z$) denotes that the generator has been trained to generate data with a high probability. In other words, the $GandD$ are the most effective at creating and distinguishing data, respectively. The ordinary GAN model is quick to train, but it has a high likelihood of divergence issues and low stability. As a result, we present a Conditional Generative Adversarial Network (CGAN). In both $GandD$, CGAN grew to the CNN model structure. This model is created by deleting the pooling layer from the CNN model and adding batch normalization between the convolutional layers. It can achieve local normalization and improve the overall network stability. After about 2000 steps, D and G reach their equilibrium, with the monitor implementation training as quickly as possible is defined in Table 2. The $DlossandGloss,$are defined in Eq. (13) and (14).

$D_{loss}=\frac{1}{m}{\displaystyle \sum }_{i=1}^m[\log D\left(x|y^{\left(i\right)}\right)+\log (1-D(G(z|y^{(i)})))]$ (13)

$G_{loss}=\frac{1}{m}{\displaystyle \sum }_{i=1}^m\log (1-D(G(z|y^{(i)})))$ (14)

The generative and discriminative models are optimized in turn to avoid over fitting on a limited dataset during the training phase. In the proposed model, we improve the discriminative model D by two steps and the generative model G by one step. According to theory, the entire adversarial process can reach Nash equilibrium if D and G are well-designed and well-taught. The distribution of manufactured data is identical to the distribution of training data in this equilibrium, and this manufactured data can be mistaken for real data due to the discriminative model's inability to distinguish between them. The Proposed Generator and discriminator GAN architecture model as shown in the Figure 3.

Table 2
Proposed GAN hyper parameters used in the generator

Generator
Operation	Kernel	Strides	Feature maps	Dropout	Nonlinearity
Gx(z) - 8 × 8 × 256 input	5X5	2X2	128	0.4	Leaky ReLU
Transposed conv.	5X5	2X2	64	0.4	Leaky ReLU
Transposed conv.	5X5	2X2	32	0.4	Leaky ReLU
Transposed conv.	5X5	2X2	1	0.4	Leaky ReLU

Conditional Generative Adversarial Networks (CGAN) Technique

The framework's optimization objective is to reduce the mean square error between the predicted sample label and the actual sample label. In order to minimize the function $\log \left(1-D(z|y);\theta G\right)$the generator is trained. As a result, the GAN's optimization problem can be stated in Eq. (15).

$Mi{n}_{G}Ma{x}_{D}V(D,G)=E_{x-p(x)}\left[\log (D\left(x|y\right));\theta G\}\right]+E_{z-p(z)}\left[\log (1-D\left(G\left(z|y\right)\right));\theta G)\right]$ (15)

Where V is the value function, and E is the predictable value. Rd is x, the random noise is z, and the distribution is P(•). In other words, the discriminator tries to tell if the generated data is real or false.

Pseudocode for Conditional Generative Adversarial Networks (CGAN) Technique:

# G e n e r a t o r N e t w o r k f u n c t i o n c r e a t e _ g e n e r a t o r ( ) : g e n e r a t o r = S e q u e n t i a l ( ) g e n e r a t o r . a d d ( D e n s e ( u n i t s = 256, i n p u t _ dim = ( l a t e n t _ dim + n u m _ c l a s s e s ) ) ) g e n e r a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) g e n e r a t o r . a d d ( B a t c h N o r m a l i z a t i o n ( ) ) g e n e r a t o r . a d d ( D e n s e ( u n i t s = 512 ) ) g e n e r a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) g e n e r a t o r . a d d ( B a t c h N o r m a l i z a t i o n ( ) ) g e n e r a t o r . a d d ( D e n s e ( u n i t s = 1024 ) ) g e n e r a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) g e n e r a t o r . a d d ( B a t c h N o r m a l i z a t i o n ( ) ) g e n e r a t o r . a d d ( D e n s e ( u n i t s = f e a t u r e s _ dim , a c t i v a t i o n = ' tanh ' ) ) r e t u r n g e n e r a t o r f u n c t i o n c r e a t e _ d i s c r i min a t o r ( ) : d i s c r i min a t o r = S e q u e n t i a l ( ) d i s c r i min a t o r . a d d ( D e n s e ( u n i t s = 1024, i n p u t _ dim = ( f e a t u r e s _ dim + n u m _ c l a s s e s ) ) ) d i s c r i min a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) d i s c r i min a t o r . a d d ( D r o p o u t ( 0.3 ) ) d i s c r i min a t o r . a d d ( D e n s e ( u n i t s = 512 ) ) d i s c r i min a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) d i s c r i min a t o r . a d d ( D r o p o u t ( 0.3 ) ) d i s c r i min a t o r . a d d ( D e n s e ( u n i t s = 256 ) ) d i s c r i min a t o r . a d d ( L e a k y Re L U ( 0.2 ) ) d i s c r i min a t o r . a d d ( D r o p o u t ( 0.3 ) ) d i s c r i min a t o r . a d d ( D e n s e ( u n i t s = 1, a c t i v a t i o n = ' s i g m o i d ' ) ) r e t u r n d i s c r i min a t o r # C G A N N e t w o r k f u n c t i o n c r e a t e _ c g a n ( g e n e r a t o r , d i s c r i min a t o r ) : c g a n _ i n p u t = I n p u t ( s h a p e = ( l a t e n t _ dim + n u m _ c l a s s e s , ) ) g e n e r a t e d _ i m a g e = g e n e r a t o r ( c g a n _ i n p u t ) c g a n _ o u t p u t = d i s c r i min a t o r ( g e n e r a t e d _ i m a g e ) c g a n = M o d e l ( c g a n _ i n p u t , c g a n _ o u t p u t ) r e t u r n c g a n # T r a i n i n g l o o p f o r e p o c h i n r a n g e ( n u m _ e p o c h s ) : f o r b a t c h i n r a n g e ( n u m _ b a t c h e s ) : # G e n e r a t e r a n d o m n o i s e a n d l a b e l s f o r t h e g e n e r a t o r i n p u t n o i s e = n p . r a n d o m . n o r m a l ( 0, 1, ( b a t c h _ s i z e , l a t e n t _ dim ) ) l a b e l s = n p . r a n d o m . r a n d int ( 0, n u m _ c l a s s e s , b a t c h _ s i z e ) . r e s h a p e ( − 1, 1 ) l a b e l s _ e n c o d e d = t o _ c a t e g o r i c a l ( l a b e l s , n u m _ c l a s s e s = n u m _ c l a s s e s ) # G e n e r a t e f a k e i m a g e s f r o m n o i s e a n d l a b e l s g e n e r a t e d _ i m a g e s = g e n e r a t o r . p r e d i c t ( n p . c o n c a t e n a t e ( ( n o i s e , l a b e l s _ e n c o d e d ) , a x i s = 1 ) ) # S e l e c t a r a n d o m b a t c h o f r e a l i m a g e s a n d c o r r e s p o n d i n g l a b e l s r e a l _ i m a g e s = ... # L o a d r e a l i m a g e s r e a l _ l a b e l s = ... # L o a d r e a l l a b e l s # C r e a t e a t r a i n i n g d a t a s e t w i t h r e a l a n d f a k e i m a g e s a n d l a b e l s c o m b i n e d _ i m a g e s = n p . c o n c a t e n a t e ( ( r e a l _ i m a g e s , g e n e r a t e d _ i m a g e s ) ) c o m b i n e d _ l a b e l s = n p . c o n c a t e n a t e ( ( r e a l _ l a b e l s , l a b e l s _ e n c o d e d ) ) # A s s i g n l a b e l s : 1 f o r r e a l i m a g e s , 0 f o r f a k e i m a g e s l a b e l s = n p . c o n c a t e n a t e ( ( n p . o n e s ( ( b a t c h _ s i z e , 1 ) ) , n p . z e r o s ( ( b a t c h _ s i z e , 1 ) ) ) ) # T r a i n t h e d i s c r i min a t o r d i s c r i min a t o r _ l o s s = d i s c r i min a t o r . t r a i n _ o n _ b a t c h ( c o m b i n e d _ i m a g e s , l a b e l s ) # T r a i n t h e g e n e r a t o r v i a t h e C G A N mod e l n o i s e = n p . r a n d o m . n o r m a l ( 0, 1, ( b a t c h _ s i z e , l a t e n t _ dim ) ) g _ l o s s = c g a n . t r a i n _ o n _ b a t c h ( n p . c o n c a t e n a t e ( ( n o i s e , l a b e l s _ e n c o d e d ) , a x i s = 1 ) , n p . o n e s ( ( b a t c h _ s i z e , 1 ) ) ) # U p d a t e p r o g r e s s a n d l o s s m e t r i c s # G e n e r a t e a n d s a v e s a m p l e i m a g e s a t t h e e n d o f e a c h e p o c h s a v e _ g e n e r a t e d _ i m a g e s ( g e n e r a t o r , e p o c h )

$\#GeneratorNetwork$ $functioncreate\_generator():$ $generator=Sequential()$ $generator.add\left(Dense\left(units=\mathrm{256,}input\_\dim =\left(latent\_\dim +num\_classes\right)\right)\right)$ $generator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $generator.add(BatchNormalization())$ $generator.add\left(Dense\left(units=512\right)\right)$ $generator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $generator.add(BatchNormalization())$ $generator.add\left(Dense\left(units=1024\right)\right)$ $generator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $generator.add(BatchNormalization())$ $generator.add\left(Dense\left(units=features\_\dim ,activation=\text{'}\tanh \text{'}\right)\right)$ $returngenerator$ $functioncreate\_discri\min ator():$ $discri\min ator=Sequential()$ $discri\min ator.add\left(Dense\left(units=\mathrm{1024,}input\_\dim =\left(features\_\dim +num\_classes\right)\right)\right)$ $discri\min ator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $discri\min ator.add\left(Dropout\left(0.3\right)\right)$ $discri\min ator.add\left(Dense\left(units=512\right)\right)$ $discri\min ator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $discri\min ator.add\left(Dropout\left(0.3\right)\right)$ $discri\min ator.add\left(Dense\left(units=256\right)\right)$ $discri\min ator.add\left(Leaky\mathrm{Re}LU\left(0.2\right)\right)$ $discri\min ator.add\left(Dropout\left(0.3\right)\right)$ $discri\min ator.add\left(Dense\left(units=\mathrm{1,}activation=\text{'}sigmoid\text{'}\right)\right)$ $returndiscri\min ator$ $\#CGANNetwork$ $functioncreate\_cgan\left(generator,discri\min ator\right):$ $cgan\_input=Input\left(shape=\left(latent\_\dim +num\_classes,\right)\right)$ $generated\_image=generator\left(cgan\_input\right)$ $cgan\_output=discri\min ator\left(generated\_image\right)$ $cgan=Model\left(cgan\_input,cgan\_output\right)$ $returncgan$ $\#Trainingloop$ $forepochinrange\left(num\_epochs\right):$ $forbatchinrange\left(num\_batches\right):$ $\#Generaterandomnoiseandlabelsforthegeneratorinput$ $noise=np.random.normal\left(\mathrm{0,}\mathrm{1,}\left(batch\_size,latent\_\dim \right)\right)$ $labels=np.random.rand\mathrm{int}\left(\mathrm{0,}num\_classes,batch\_size\right).reshape\left(-\mathrm{1,}1\right)$ $labels\_encoded=to\_categorical\left(labels,num\_classes=num\_classes\right)$ $\#Generatefakeimagesfromnoiseandlabels$ $generated\_images=generator.predict\left(np.concatenate\left(\left(noise,labels\_encoded\right),axis=1\right)\right)$ $\#Selectarandombatchofrealimagesandcorrespondinglabels$ $real\_images=\mathrm{...}\#Loadrealimages$ $real\_labels=\mathrm{...}\#Loadreallabels$ $\#Createatrainingdatasetwithrealandfakeimagesandlabels$ $combined\_images=np.concatenate\left(\left(real\_images,generated\_images\right)\right)$ $combined\_labels=np.concatenate\left(\left(real\_labels,labels\_encoded\right)\right)$ $\#Assignlabels:1forrealimages,0forfakeimages$ $labels=np.concatenate\left(\left(np.ones\left(\left(batch\_size,1\right)\right),np.zeros\left(\left(batch\_size,1\right)\right)\right)\right)$ $\#Trainthediscri\min ator$ $discri\min ator\_loss=discri\min ator.train\_on\_batch\left(combined\_images,labels\right)$ $\#TrainthegeneratorviatheCGAN\mathrm{mod}el$ $noise=np.random.normal\left(\mathrm{0,}\mathrm{1,}\left(batch\_size,latent\_\dim \right)\right)$ $g\_loss=cgan.train\_on\_batch\left(np.concatenate\left(\left(noise,labels\_encoded\right),axis=1\right),np.ones\left(\left(batch\_size,1\right)\right)\right)$ $\#Updateprogressandlossmetrics$ $\#Generateandsavesampleimagesattheendofeachepoch$ $save\_generated\_images\left(generator,epoch\right)$

As a result, the loss in this binary classification is determined by cross-entropyin Eq. (16).

$Los{s}_D=\frac{1}{n}{\displaystyle \sum }_{i=1}^n{y}_i\log D\left(x_i\right)-\frac{1}{n}{\displaystyle \sum }_{i=1}^n(1-y_i)\log (1-D\left(x_i\right))$ (16)

The goal of GAN training is to discover the Nash equilibrium of a non-convex problem with continuous, high-dimensional parameters. To find the lowest value of a cost function, GANs are generally trained using gradient descent methods. The GAN learns the feature representation in the absence of a cost function. Table 3 depicts the parameters of both the generator and discriminator models in the Proposed GAN hyperparameters.

Table 3
Projected GAN hyper parameters are in Discriminator.

Operation	Kernel	Strides	Feature maps	Dropout	Nonlinearity
Gx(z) - 32 × 32 × n input Transposed conv.	5X5	2X2	64	0.4	Leaky ReLU
Transposed conv.	5X5	2X2	128	0.4	Leaky ReLU
Transposed conv.	5X5	2X2	256	0.4	Leaky ReLU
Transposed conv.	5X5	1X1	512	0.4	Sigmoid

Proposed Scalp swarm optimization (SSA) algorithm

To increase the classification accuracy of an EEG signal using the created fake EEG data, the generated data must include label info. As a result, the above-mentioned CGAN is expanded to improve the classifier's learning rate using the Scalp swarm optimization method. The Scalp swarm algorithm (SSA) is a novel optimization approach developed to tackle a wide range of optimization issues. It resembles the activity of Scalps in nature; Scalps are barrel-shaped planktonic tunicates from the Scalpidae family. Furthermore, they have tissues that are comparable to jellyfishes, as well as movement activity and a high water percentage in their weight. SSA begins by categorizing the population into two groups: leaders and followers. The front scalp of the chain is referred to as the leader, while the other Scalps are referred to as the followers. The Scalps location is decided in n dimensions, which reflect a problem's search space, and n represents the variables in the problem. These Scalps look for a food supply, which signals the swarm's objective [²²]. Because the position must be updated on a regular basis, the following equation is utilized to conduct this operation with the Scalp leader in Eq. (17).

$x_j^i=\{\begin{array}{c}{F}_j+c_1\left(\left(u{b}_j-l{b}_j\right)*c_2+l{b}_j\right)C_3\le 0\\ F_j-c_1\left(\left(u{b}_j-l{b}_j\right)*c_2+l{b}_j\right)C_3>0\end{array}$ (17)

Where$x_j^i$ is the site of the leader within jth dimension, where the food source in this dimension is $F_j$, the upper and the lower bounds are$u{b}_j$ and$l{b}_j$, respectively. $c_2$and $c_3$ are generated arbitrarily in the range [0, 1] to keep the search area clean As a result of its role in balancing the exploration and exploitation phases, this algorithm's parameter $c_1$ is an extremely important coefficient in Eq. (18).

$c_1=2e^{-\left(\frac{4t}{t_{\max }}\right)}$ (18)

Where $t$ and $t_{\max }$ specify the sum of iterations currently being performed and the maximum number that can be performed. In order to update the position of the followers, the SSA first has to update the leader's position in Eq. (19):

$x_j^i=\frac{1}{2}(x_j^i+x_j^{i-1})$ (19)

where $x_j^i$ is defined as the $i-th$ follower location within $j-th$ dimension and$i$ is better than 1.

RESULTS

MATLAB software is used for algorithm analysis and data sampling. In this experimental analysis, we have used the BCI IV a dataset is a challenge dataset, and there is a protocol to use it, in this dataset we have evaluated the dataset into two parts, in first 70% of dataset for training and remaining 30% for testing and validation. After the classification algorithm parameters are modified by optimization technique. One of the most well-known tools for computer vision is the MATLAB software. The suggested algorithms are scripted and function-programmed. Validation measures three common criteria, including sensitivity, accuracy, and precision. In this experimental study, we mainly comparing the proposed CGAN with CNN classifier on both AlexNet and Resnet models, these both model hyper parameters as in Alexnet model, used the input layer size of [128 128], also used the three 2D convolution layer, three batch Normalization Layer, three relu Layer, one softmax layer and fully Connected Layer. The training option is conducted by using stochastic gradient descent with momentum (SGDM) as a popular optimization algorithm, which accelerate the gradients vectors in the right direction, thus leading to faster converging. In Resnet model, the Input Layer size of [32 32 1]. Here, the convolution 2d Layer with to ('padding','same'), three batch Normalization Layers, three relu Layer, and one fully Connected Layer (4), softmax Layer and classification Layer with training option of SGDM was used.

The evaluation metrics include sensitivity, specificity, precision and accuracy, which are used to analysis the performance of the proposed model [23]. In qualitative analysis, we assess the performance of the proposed technique under different techniques with different class as two-class and multi-class performance. The classification techniques as CNN (ResNet and AlexNet) and GAN model. Moreover, we applied different optimization techniques as PSO, ACO, whale optimization, and Scalp swarm optimization techniques, which methods are applied in both classes and each subject in the class [²⁴]. Table 4 shows the confusion matrix performance of the proposed Scalp Swam_EMD_DWT_GAN (Multi class). Table 5 shows the confusion matrix performance of the proposed Scalp Swam_EMD_DWT_GAN (Two classes).

Table 4
Confusion matrix for proposed ScalpSwam_EMD_DWT_CGAN (Multi class)

Actual Classes
	Left hand	Right hand	Foot	Tongue
Left hand	88.31	5.13	05.01	01.55
Right hand	01.30	87.21	07.54	03.95
Foot	06.54average	01.39	89.32	02.75
Tongue	01.52	04.63	06.82	87.03

Table 5
Confusion matrix for proposed ScalpSwam_EMD_DWT_GAN (Two classes)

Predicted Class	Actual Classes
	Left hand	Left hand	Right hand
	Right hand	89.64	10.36
		10.14	89.86

Table 6. displays the performance analysis of CNN (AlexNet) model with different optimization techniques by using two classes. In this analysis, we evaluated performance by using different optimization techniques, Scalp swarm optimization technique performs the better classification performance than other optimization techniques. Figure 4. illustrate the one-task average feature power of the real and generated EEG in accordance with the electrodes' placements on topographic maps to further explain the spatial feature extraction.

Table 6
Performance analysis of CNN (AlexNet) model with different optimization techniques by using BCI IV (a) (two classes)

Subjects	Performance measures	ScalpSwam_EMD_DWT _CNN (AlexNet)	Whale_EMD_DWT_CNN (AlexNet)	ACO_EMD_DWT_CNN (AlexNet)	PSO_EMD_DWT_CNN (AlexNet)
Al	Precision	87.56	88.50	84.51	84.59
	Sensitivity	88.42	88.54	84.53	84.62
	Specificity	87.65	88.56	84.56	84.54
	Accuracy	88.00	89.88	84.02	84.09
Aa	Precision	84.98	90.54	87.22	86.54
	Sensitivity	98.57	90.00	87.46	86.54
	Specificity	89.55	90.43	87.99	86.52
	Accuracy	90.09	90.65	87.90	86.68
Ay	Precision	89.39	85.32	77.91	81.45
	Sensitivity	89.75	85.85	77.84	81.56
	Specificity	89.53	85.41	77.89	81.51
	Accuracy	90.98	85.03	77.99	81.89
Av	Precision	92.56	86.53	84.09	83.39
	Sensitivity	92.65	86.65	84.02	83.86
	Specificity	92.85	86.95	84.62	83.39
	Accuracy	92.00	86.33	84.95	83.45
Aw	Precision	89.66	85.39	82.84	82.48
	Sensitivity	89.52	85.98	82.54	82.29
	Specificity	89.85	85.35	82.00	82.69
	Accuracy	87.85	85.68	82.36	82.85

Figure 4
The average feature power on topographic maps

It's critical to have training stability when using GANs. The loss suffered by the network during the training iterations is an excellent predictor of how well the training is going. For one randomly selected individual, the generator and discriminator losses are shown in Figure 5 across 1000 iterations. The generator loss should gradually decrease and both networks should reach constant values after successful training. After around 250 iterations, the generator loss steadily diminishes and both losses return to their initial values.

Figure 5
Generator and discriminator losses for 1000 epoches

Table 7 shows that the performance analysis of CNN (ResNet) and GAN model with different optimization techniques by using two classes. In this comparison analysis, using the various optimization techniques for CNN (ResNet) and GAN model, the ScalpSwam_EMD_DWT_GAN model reached better results than the other models, the better classification accuracy of the proposed model as 97.00%, which is obtained in “aa” subject.

Table 7
Performance analysis of CNN (ResNet) and GAN model with different optimization techniques by using BCI IV (a) (two classes)

Subjects	Performance measures	ScalpSwam_ EMD_DWT_CNN (ResNet)	Whale_EMD_ DWT_CNN (ResNet)	ACO_EMD_ DWT_CNN (ResNet)	PSO_EMD_DWT_ CNN(ResNet)	ScalpSwam_EMD_ DWT_CGAN
Al	Precision	90.54	90.54	89.75	86.33	91.20
	Sensitivity	90.65	90.54	89.51	86.64	90.69
	Specificity	89.25	90.15	89.52	86.25	91.30
	Accuracy	90.33	89.33	90.00	86.14	90.05
Aa	Precision	91.55	88.52	90.53	87.55	91.55
	Sensitivity	90.59	91.75	90.54	87.59	90.56
	Specificity	91.54	90.52	90.52	87.52	90.53
	Accuracy	91.39	90.32	90.66	88.66	91.96
Ay	Precision	89.57	88.25	85.53	82.50	90.82
	Sensitivity	90.89	88.85	85.53	82.02	91.84
	Specificity	91.11	88.65	85.54	82.14	90.85
	Accuracy	90.63	88.14	85.20	82.77	90.98
Aw	Precision	91.55	87.50	86.54	83.97	90.15
	Sensitivity	89.54	87.50	86.56	83.49	90.45
	Specificity	91.51	87.50	86.56	83.67	90.54
	Accuracy	90.00	88.20	86.21	83.63	91.98
Av	Precision	91.34	89.54	87.52	82.54	91.53
	Sensitivity	94.22	89.54	87.51	82.51	90.51
	Specificity	91.19	89.12	87.59	82.52	91.50
	Accuracy	89.78	89.95	88.25	82.33	91.09

Table 8 shows that the CNN (AlexNet) performance of the model with different optimization techniques by using BCI III (a) (multi-class). The results obtained by using different optimization techniques, by this comparison analysis, Scalp swims optimization provide better classification results than other optimization models.

Table 8
Performance analysis of CNN (AlexNet) model with different optimization techniques by using BCI III (a) (multi-class)

Subjects	Performance measures	ScalpSwam_EMD_DWT_CNN (AlexNet)	Whale_EMD_DWT_CNN (AlexNet)	ACO_EMD_DWT_CNN (AlexNet)	PSO_EMD_DWT_CNN (AlexNet)
Multi class k3b	Precision	81.25	93.53	81.25	78.42
	Sensitivity	85.10	82.53	81.41	77.51
	Specificity	90.89	87.87	93.54	91.53
	Accuracy	85.62	82.87	81.00	78.23
Multi class k6b	Precision	79.66	79.75	75.57	74.25
	Sensitivity	79.25	79.25	74.75	74.25
	Specificity	93.46	93.33	91.75	97.25
	Accuracy	79.45	79.88	75.17	74.13
Multi class L1b	Precision	85.25	83.21	81.25	78.89
	Sensitivity	85.82	82.50	81.16	77.50
	Specificity	95.00	94.31	93.50	92.50
	Accuracy	85.91	82.70	81.12	78.87

Table 9
Performance analysis of CNN (ResNet) and GAN model with different optimization techniques by using BCI III (a) (multi-class)

Subjects	Performance measures	ScalpSwam_ EMD_DWT_ CNN (ResNet)	Whale_EMD_DWT_CNN (ResNet)	ACO_EMD_DWT_CNN (ResNet)	PSO_EMD_DWT_CNN (ResNet)	ScalpSwam_EMD_DWT_DCGAN (proposed)
Multi class k3b	Precision	90.02	87.42	87.75	83.50	91.51
	Sensitivity	90.33	86.75	87.52	83.25	91.45
	Specificity	97.56	95.75	96.58	89.51	97.32
	Accuracy	90.95	87.26	88.45	83.00	91.11
Multi class k6b	Precision	86.25	84.53	84.52	78.75	90.75
	Sensitivity	87.75	84.53	84.25	80.51	90.75
	Specificity	96.64	94.75	94.75	92.75	91.25
	Accuracy	88.49	84.03	84.45	78.85	91.09
Multi class L1b	Precision	90.12	87.99	87.75	83.51	93.62
	Sensitivity	90.36	86.75	87.52	83.25	90.15
	Specificity	97.86	87.50	85.21	88.50	91.35
	Accuracy	90.46	87.45	88.33	83.00	91.18

Table 9 represents the performance measure of CNN (ResNet) and GAN model with different optimization techniques by using BCI III (a) (multi-class). In multiclass it has three subject as K3b, K6b and L1b. In this each subject, performance is evaluated by using four parametric metrics such as precision, Sensitivity, specificity and Accuracy. The performance comparisons of the proposed model reached the better performance than other techniques.

Figures 6 and 7 show that the graphical representation of the proposed technique with CNN, AlexNet, and ResNet combination model accuracy comparisons of both classes and multi-class. In the two classes, proposed attained the better performance than multi-class. In two class “aa” subjects, the proposed model reaches the better accuracy of 97.96%. In the multiclass K3b subjects, the proposed model reaches the highest accuracy of 94.11%, respectively.

Figure 6
Graphical representation of the accuracy performance of CNN_ Alexnet combinations

Figure 7
Graphical representation of the accuracy performance of CNN_ ResNet model with the proposed method

DISCUSSION

Table 10 shows the quantitative analysis of the proposed with existing model comparison. In this comparison, we compared some existing techniques, which used different machine learning and deep learning classifier models and some extraction techniques.

Table 10
Comparison performance analysis of the proposed with existing methods.

Reference	Brain signal type	Electrode placement	Feature Extraction	Classification	Optimization	Accuracy (%)
[25]	Motor Imagery	C3, C4 and Cz	Regularized commonspatial patterns	SVM	-	85.10
[26]	Motor Imagery	18 Electrodes	CSP-AM	SVM	Bat Optimization	86.60
[27]	Motor Imagery	C3, C4 and Cz	SFB-CSP	SVM	-	82.56
[28]	Motor Imagery	C3, C4 and Cz	DWT	ANN	-	84.77
[28]	Motor Imagery	C3, C4 and Cz	EMD	ANN	-	90.02
[29]	Motor Imagery	C3, C4 and CZ	EMD	CNN (2 Class)	-	89.38
[29]	Motor Imagery	C3, C4and CZ	EMD	CGAN(2Class)	-	95.29
[30]		C3, C4and CZ	-	Attention Temporal CNN		81.71
Proposed (4-class)	Motor Imagery	C3, C4 and Cz	EMD -DWT	CNN	Scalp swarm	81.61
Proposed (4-class)	Motor Imagery	C3, C4 and Cz	EMD -DWT	CGAN	Scalp swarm	89.11

The listed existing models only employed extraction and classification techniques; they left out optimization techniques, which reduced the classification performance. In contrast, the proposed model applied the Scalp swarm optimization technique to get over the earlier restrictions and accuracy degradation. By this comparison, our proposed model reached the better classification accuracy in both two and multi-class classification. The restrictions of the proposed method with respect to the inaccuracy of results. The anatomy of the human brain is amazing. Even as human beings, our thoughts occasionally leave us perplexed. As a result, it is unrealistic to anticipate that a BCI will be able to discern every brain impulse. BCI occasionally may misinterpret the user's intentions. A person utilizing prosthesis to raise his index finger may be mistakenly identified by the BCI, resulting in the middle finger being lifted instead. The possibility of inaccurate results when using BCI technology is a serious concern. The security of your data is not guaranteed by BCI technology, on the other hand. The automated device allows anyone to read your thoughts. An adversary nation's hackers might potentially access military personnel's minds when BCI is used for military operations and eventually divulge all secret data. In the next five years, assailants might be able to change victims' memories.

CONCLUSION

The classification performance improved in MI tasks we used a GAN model-based DL strategy in this experiment. Empirical mode decomposition is used to recreate the MI signal (EMD), which can manage the non-stationary problem and mix their intrinsic mode functions (IMFs). The accuracy of our classification was tested on two datasets from the BCI competitions III and IV. Two and multiclass evaluations are used in this case to recognize the EEG signal. The multiclass evaluation the classifier performs goes down, and thus planned to implement the optimization techniques as Scalp swarm for classifier model. The optimization can improve the learning rate and optimize the parameters of the classifier, which is most helpful for attaining better results in BCI scheme. However, the proposed technique significance is estimated and compared with different methodologies. The proposed classifier and optimization model proved to produce the better performance compared with existing models. The performance accuracy of the proposed model is 81.61% for CNN multi class and 89.11% for CGAN multi class. Furthermore, BCI systems will deeply rely on the computing environment, like BCI systems on embedded platform and application domain. It will design and implementation of a BCI system based on embedded system, which will enable humans to command artificial peripherals by simply thinking about the task efficiently.

REFERENCES

• ¹ Aggarwal S, Chugh N. Signal processing techniques for motor imagery brain computer interface: A review Array. 2020 Apr; 1:100003. https://doi.org/10.1016/j.array.2019.100003
  » https://doi.org/10.1016/j.array.2019.100003
• ² Hang W, Feng W, Liang S, Wang Q, Liu X, Kup Sze C. Deep stacked support matrix machine-based representation learning for motor imagery EEG classification. Computer methods and programs in biomedicine.2020 Sep;193:105466. https://doi.org/10.1016/j.cmpb.2020.105466
  » https://doi.org/10.1016/j.cmpb.2020.105466
• ³ Li MA, Han JF, Duan LJ. A novel MI-EEG imaging with the location information of electrodes. IEEE Access 2019 Dec;8:3197-211. https://doi.org/10.1109/ACCESS.2019.2962740
  » https://doi.org/10.1109/ACCESS.2019.2962740
• ⁴ Khare SK, Bajaj V. A facile and flexible motor imagery classification using electroencephalogram signals. Computer Methods and Programs in Biomedicine.2020 Dec;197:105722. https://doi.org/10.1016/j.cmpb.2020.105722
  » https://doi.org/10.1016/j.cmpb.2020.105722
• ⁵ Mandeep KG, Satvir S. Investigation of EEG signal classification techniques for Brain Computer Interface. Int. J. Eng. Adv. Technol. 2020 Feb;9(3):3266-74.
• ⁶ Ogudo KA, Surendran R, Osamah Ibrahim K. Optimal Artificial Intelligence-Based Automated Skin Lesion Detection and Classification Model. Computer Systems Science and Engineering, 2022 Jun;44(1),693-707. https://doi.org/10.32604/csse.2023.024154
  » https://doi.org/10.32604/csse.2023.024154
• ⁷ Vijayasankar A, Kumar PR. Correction of blink artifacts from single channel EEG by EMD-IMF thresholding. In 2018 Conference on Signal Processing and Communication Engineering Systems (SPACES). 2018 Mar;176-80. https://doi.org/10.1109/SPACES.2018.8316340
  » https://doi.org/10.1109/SPACES.2018.8316340
• ⁸ Tariq M, Trivailo PM, Simic M. Classification of left and right knee extension motor imagery using common spatial pattern for BCI applications. Procedia Comput. Sci. 2019;159:2598-606. https://doi.org/10.1016/j.procs.2019.09.256
  » https://doi.org/10.1016/j.procs.2019.09.256
• ⁹ Dai M, Zheng D, Liu S, Zhang P. Transfer kernel common spatial patterns for motor imagery brain-computer interface classification. Comput Math. Methods Med.2018 Feb, 9871603.https://doi.org/10.1155/2018/9871603
  » https://doi.org/10.1155/2018/9871603
• ¹⁰ Alotaibi Y, Noman Malik M, Hayat Khan H, Batool A, UlIslam S, Alsufyani A, et al . Suggestion mining from opinionated text of big social media data. Computers, Materials & Continua. 2021 May;68(3):3323-38. https://doi.org/10.32604/cmc.2021.016727
  » https://doi.org/10.32604/cmc.2021.016727
• ¹¹ Fahimi F, Zhang Z, Goh WB, Ang KK, Guan C. Towards EEG generation using GANs for BCI applications. 2019 IEEE EMBS Int. Conf. Biomed. Health Informat. (BHI). 2019 May;1-4. https://doi.org/10.1109/BHI.2019.8834503
  » https://doi.org/10.1109/BHI.2019.8834503
• ¹² Corley IA, Huang Y. Deep EEG super-resolution: Upsampling EEG spatial resolution with generative adversarial networks. In Proc. IEEE EMBS Int. Conf. Biomed and Health Informat. (BHI). 2018 Mar;100-3. http://dx.doi.org/10.1109/BHI.2018.8333379
  » http://dx.doi.org/10.1109/BHI.2018.8333379
• ¹³ Ramadan RA, Vasilakos AV. Brain computer interface: control signals review. Neurocomputing. 2016 Oct; 223:26-44. http://dx.doi.org/10.1016/j.neucom.2016.10.024
  » http://dx.doi.org/10.1016/j.neucom.2016.10.024
• ¹⁴ Islam MR, Rahim MA, Akter H, Kabir R, Shi J. Optimal IMF selection of EMD for sleep disorder diagnosis using EEG signals. In Proceedings of the 3rd International Conference on Applications in Information Technology. 2018 Nov;96-101. https://doi.org/10.1145/3274856.3274876
  » https://doi.org/10.1145/3274856.3274876
• ¹⁵ Hazra D, Byun YC. SynSigGAN: Generative adversarial networks for synthetic biomedical signal generation. Biology. 2020 Dec;9(12):441. https://doi.org/10.3390/biology9120441
  » https://doi.org/10.3390/biology9120441
• ¹⁶ Roy S, Dora S, McCreadie K, Prasad G. MIEEG-GAN: generating artificial motor imagery electroencephalography signals. In 2020 International Joint Conference on Neural Networks (IJCNN), IEEE. 2020 Mar;1-8.
• ¹⁷ Mao, X. Li Q, Xie H, Raymond YK, Lau, Wang Z, Smolley SP,et al. Least squares generative adversarial networks. Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2017;2794-802. https://arxiv.org/abs/1611.04076
  » https://arxiv.org/abs/1611.04076
• ¹⁸ Rehman SU, Tu S, Rehman OU, Huang Y, Optimization of CNN through a novel training strategy for visual classification problems. Entropy, 2018 Apr;20(4):290. https://doi.org/10.3390/e20040290
  » https://doi.org/10.3390/e20040290
• ¹⁹ Dagdevir E, Tokmakci M. Optimization of preprocessing stage in EEG based BCI systems in terms of accuracy and timing cost. Biomedical Signal Processing and Control.2021 May;67: 102548. https://doi.org/10.1016/j.bspc.2021.102548
  » https://doi.org/10.1016/j.bspc.2021.102548
• ²⁰ Yuanhao C. Image and Graph Restoration Dependent on Generative Adversarial Network Algorithm. Technical Gazette. 2021;28(6):1820-24. https://doi.org/10.17559/TV-20210413111850
  » https://doi.org/10.17559/TV-20210413111850
• ²¹ Srirangam, S. Scalp Swarm Optimization Based Machine Learning Algorithm for Epileptic Seizure Detection and Classification Model. Turk. J. Comput. Math. Educ. (TURCOMAT). 2021;12(14):799-808.
• ²² Venkatachalam K, Devipriya A, Maniraj J, Sivaram M, Ambikapathy A, Iraj SA, et al. A Novel Method of motor imagery classification using EEG signal. Artificial intelligence in medicine. 2020 Mar;103:101787. https://doi.org/10.1016/j.artmed.2019.101787
  » https://doi.org/10.1016/j.artmed.2019.101787
• ²³ AlSaegh A, Dawwd SA, AbdulJabbar JM. CutCat: An augmentation method for EEG classification. Neural Networks. 2021 Sep;141:433-43. https://doi.org/10.1016/j.neunet.2021.05.032
  » https://doi.org/10.1016/j.neunet.2021.05.032
• ²⁴ Jin J, Miao Y, Daly I, Zuo C, Hu D, Cichocki A, et al. Correlation-based channel selection and regularized feature optimization for MI-based BCI. Neural Networks, 2019 Oct;118:262-70. https://doi.org/10.1016/j.neunet.2019.07.008
  » https://doi.org/10.1016/j.neunet.2019.07.008
• ²⁵ Selim S, Tantawi MM, Shedeed HA, Badr A. A csp\am-ba-svm approach for motor imagery bci system. IEEE Access. 2018;6:49192-208. https://doi.org/10.1109/ACCESS.2018.2868178
  » https://doi.org/10.1109/ACCESS.2018.2868178
• ²⁶ Stephe S, Jayasankar T, Vinoth Kumar K. Motor Imagery Recognition of EEG Signal using Cuckoo-Search Masking Empirical Mode Decomposition. Int. J. Innov. Technol. Explor. Eng. 2019;8(11):2717-20. https://doi.org/10.35940/ijitee.K2175.0981119
  » https://doi.org/10.35940/ijitee.K2175.0981119
• ²⁷ Abdalsalam E, Yusoff MZ, Mahmoud D, Malik A, Bahloul MR. Discrimination of four classes simple limb motor imagery movements by brain-computer interface. Biomed. Signal Process. Control. 2018 May;44:181-90. http://dx.doi.org/10.1016/j.bspc.2018.04.010
  » http://dx.doi.org/10.1016/j.bspc.2018.04.010
• ²⁸ Stephe S, Jayasankar T, Vinoth Kumar K. Motor Imagery EEG Recognition using Deep Generative Adversarial Network with EMD for BCI Applications. Technical Gazette. 2022;29(1):92-100. https://doi.org/10.17559/TV-20210121112228
  » https://doi.org/10.17559/TV-20210121112228
• ²⁹ Mani V, Manickam P, Alotaibi Y, Alghamdi S, Khalaf OI. Hyperledger healthchain: Patient-centric IPFS-based storage of health records. Electronics. 2021;10(23):3003. https://doi.org/10.3390/electronics10233003
  » https://doi.org/10.3390/electronics10233003
• ³⁰ Altaheri H, Muhammad G, Alsulaiman M. Physics-informed attention temporal convolutional network for EEG-based motor imagery classification. IEEE Transactions on Industrial Informatics. 2022 Aug 9;19(2):2249-58. DOI: 10.1109/TII.2022.3197419
  » https://doi.org/10.1109/TII.2022.3197419

• Funding:
  The authors declare they have no funding applied.
• Data Availability Statement:
  Available Based on Request. The datasets generated and or analyzed during the current study are not publicly available due to the extension of the submitted research work. They are available from the corresponding author upon reasonable request. Proposed method combines (SSO) and (GAN) techniques for the recognition of motor imagery. While this approach has its merits, it is important to consider some potential limitations are success of any machine learning method, including GAN-based approaches, heavily relies on the availability and quality of training data. If the dataset used for training is limited in size or not representative of the target population, it may lead to suboptimal performance or generalization issues.
• Problem Statement:
  Motor imagery recognition from EEG signals is a fundamental component of brain-computer interfaces (BCIs) and holds significant potential for applications in assistive technology and healthcare. However, current methods for motor imagery EEG recognition often face challenges in achieving high accuracy and robustness. Additionally, optimizing the feature extraction process is essential for enhancing the performance of EEG-based BCIs. Scalp Swarm Optimization (SSO) is a promising optimization algorithm for feature selection, but its effectiveness and efficiency can be further improved.

Edited by

• Editor-in-Chief:
  Alexandre Rasi Aoki

• Associate Editor:
  Raja Soosaimarian Peter Raj

Data availability

Available Based on Request. The datasets generated and or analyzed during the current study are not publicly available due to the extension of the submitted research work. They are available from the corresponding author upon reasonable request. Proposed method combines (SSO) and (GAN) techniques for the recognition of motor imagery. While this approach has its merits, it is important to consider some potential limitations are success of any machine learning method, including GAN-based approaches, heavily relies on the availability and quality of training data. If the dataset used for training is limited in size or not representative of the target population, it may lead to suboptimal performance or generalization issues.

Publication Dates

• Publication in this collection
  08 Nov 2024
• Date of issue
  2024

History

• Received
  28 May 2023
• Accepted
  01 Feb 2024