Kannika Parameshwari Chinnadurai*
| Batri Krishnan | Aravind Britto Karupanan Raju
© 2025 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
Epilepsy occurs due to abnormal brain activity causing seizures. Brain injuries, trauma, or even genetic issues cause this problem resulting in physical impairments and memory-related issues. Electroencephalography (EEG) based epilepsy detection and diagnosis is eased using computer and artificial intelligence algorithms in the modern health sciences. This article thus introduces an Amplitude Differentiation Model (ADM) using Deep Learning (DL) to identify slow signals that are identified as the initial seizures. The deep learning process differentiates between successive oscillations to correlate with the training data in detecting low signal outputs. The low signal outputs are verified for their normal amplitude using DL under abnormal or normal classification. The unrestored differentiation amplitude phase represents a seizure identified using the maximum correlation factor. Both the original and unrestored oscillations are used for training the DL to improve the precision between successive amplitude changes in an EEG signal. This proposed method shows the highest level of differentiation based on observation time, where the high precision unrestored oscillation enhances accuracy with reduced correlation time. The proposed AD model increases accuracy by 15.55%, precision by 14.35%, and correlation factor by 16.08%. This model reduces correlation time by 9.55% and differentiation error by 10.78% for the various amplitude changes.
amplitude differentiation, deep learning, EEG, epilepsy, oscillation changes
Detecting epilepsy through EEG signals involves analyzing abnormal brain activities associated with seizures. Advanced algorithms, like deep learning models, enhance the identification of initial seizures by recognizing patterns in slow signals [1]. Successive oscillations are differentiated by these models, aiding in the classification of EEG outputs as abnormal or normal [2]. The precise location of seizures is determined by observing the unrestored amplitude phase using the maximum correlation factor. Continuous monitoring with wearable EEG devices holds promise for the timely detection of epilepsy events [3]. Machine learning approaches, extending beyond deep learning, contribute significantly to refining the accuracy of epilepsy diagnosis [4]. Integrating clinical data, such as patient history, further enhances the overall effectiveness of EEG-based diagnosis. Addressing challenges in EEG data preprocessing is essential for improving data quality and optimizing model performance [5].
The early signs of epilepsy depend on the EEG signals utilizing amplitude differentiation. The emphasis on amplitude differentiation facilitates the identification of specific characteristics associated with the initial phases of epileptic activity [6]. Continuous monitoring of EEG signals employing this approach presents the opportunity for timely intervention, preventing the escalation to a full-blown seizure [7]. The utilization of EEG amplitude differentiation not only enhances the precision of early epilepsy detection but also empowers models to learn and recognize distinct patterns indicative of imminent seizures [8]. It exhibits promise in advancing overall diagnostic effectiveness, offering a proactive and personalized strategy for identifying potential risks at an early stage [9]. Highlighting ethical considerations, encompassing patient consent and data privacy, in the responsible application of EEG amplitude differentiation for early epilepsy detection [10]. Further exploration of how EEG amplitude differentiation compares with alternative diagnostic methods enhances understanding of its relative effectiveness in characterizing epilepsy [11].
The application of deep learning for epilepsy detection through EEG signals involves deploying advanced neural network architectures to analyze intricate brainwave patterns. These models, trained on extensive datasets, autonomously extract features linked to epileptic activities, which significantly improve the accuracy of detection algorithms. These findings are supported by references [12, 13]. The adaptability of deep learning, which allows the identification of diverse seizure types, demonstrates its effectiveness in real-world scenarios [14]. Continuous refinement through training ensures these models remain attuned to evolving patterns in EEG data, ultimately optimizing performance over time [15]. Ethical considerations, such as transparent communication with patients, safeguarding data privacy, and obtaining informed consent, play a critical role in responsibly applying deep learning algorithms in healthcare settings [16]. Comparative analyses between deep learning-based approaches and conventional methods offer insights into advancements and potential challenges in integrating artificial intelligence for epilepsy detection [17]. The research contributions are:
The article is organized as follows: Section 2 presents the related works from different authors with their novel techniques and methods. In Section 3 the proposed model is briefed with suitable illustrations and derivations. Section 4 presents the experimental and comparative analysis using dataset and metrics respectively. Section 5 concludes the article with the findings, limitations, and future scope.
Tawhid et al. [18] presented a new neural network using convolutional long short-term memory to identify epilepsy from EEG. De Brabandere et al. [19] created a method combining automatic and hand-crafted features from EEG for improved epileptic episode identification. The goal was to make spotting seizures easier by assessing how well automatic and hand-crafted features could work together. Pandya et al. [20] introduced a method using unique features from EEG recordings for epilepsy prediction through machine learning.
Lee et al. [21] designed a quick system using a special chip and a CNN algorithm to detect epilepsy in real time. Wang et al. [22] developed a method using a mix of SVM and kernel functions to automatically spot epilepsy in EEG recordings. Shen et al. [23] introduced a method for real-time epileptic seizure detection based on EEG. The primary goal was to detect epilepsy seizures in real time using EEG data. Shoji et al. [24] created mEEGNet, a small neural network, to find issues in EEG recordings of epilepsy patients.
Chakraborty and Mitra [25] suggested a method to detect epilepsy seizures using VMD and a kurtosis-based approach. Cao et al. [26] developed a method to identify epilepsy using EEG data from seizure-free moments and machine learning. Lebal et al. [27] developed Epilepsy-Net, a model using a 1D-inception network with attention to identifying epilepsy from EEG recordings.
Majzoub et al. [28] proposed an AlexNet-based model for detecting epilepsy from multi-channel EEG signals. Goel et al. [29] created an automated method using recurrence plots and transfer learning to extract features for identifying epileptic EEG data. The main goal was to provide a dependable way to spot and categorize epileptic seizures in EEG signals. Rajinikanth et al. [30] suggested a method to identify epilepsy in EEG recordings using SET. Sunaryono et al. [31] introduced a method to automatically detect epilepsy from EEG data using wavelet-based gradient boosting machines fusion. Pandey et al. [32] developed a smart system using a mix of CNN and LSTM to find epileptic seizures in EEG signals. This article addressed the problem of low and high feature differentiation of EEG inputs to retain high analysis precision.
The research gap identified is the phase classification for amplitude differentiation observed in consecutive trials. The above-discussed methods handle phase amplitude and variations with less concern, resulting in retained detection accuracy but decreased precision in low amplitude phases. Hence, the proposed ADM also considers low-range classifications as a pre-classified model to maximize the detection range. The proposed model differs from existing methods by identifying various variations between alternate sequences to enhance the correlation rate.
The EEG signal is used to detect epilepsy in the early stage of diagnosis. Here, the ADM is introduced in this approach and estimates the better identification phase. The amplitude differentiation ranges from low too high to estimate the normal and abnormal detection of epilepsy. Deep learning is introduced in this work during the computation step for disease identification using the EEG signal. In this stage, the diagnosis is carried out on the EEG signal for signal processing that varies. At this stage, the analysis is conducted to identify epilepsy. This work associates both the characteristics and patterns of EEG signals with the electrical activity of detection. Figure 1 illustrates the proposed model.
Figure 1. Proposed ADM using DL
The computation process is utilized for deploying early-stage Epilepsy detection. Abnormal brain activity is detected in this case to identify Epilepsy from the EEG signal. Introducing Deep learning in the Amplitude differential model analyzes the differentiation of this signal. The evaluation of Epilepsy detection aims to identify seizures in this phase. The proposed work focuses on achieving high precision and accuracy while reducing the correlation time. The preliminary stage involves analyzing the EEG signal using the equation below:
\left.\beta=\begin{array}{c}\sum_{e_y}^{g_d}\left(o_e * h_c\right)+\binom{e_y / p^{\prime}+o_e}{y_c / \prod_{h_c}\left(g_d\right)} \\ * \prod_{y_c}^{o_e}\left[\begin{array}{l}\left(h_c * \frac{g_d}{e_y}\right) \\ +\left(p^{\prime} * y_c\right)\end{array}\right] \\ =\left(\frac{e_y}{\prod_{o_e}\left(g_d+h_c\right)}\right) * \sum_{p^{\prime}}\left(y_c * g_d\right)+\left(\frac{y_c * o_e}{p^{\prime} / e_y}\right)\end{array}\right\} (1)
The analysis is carried out in the above equation where the EEG signal is the input and it is represented as \beta. Epilepsy detection is observed in this equation and it is labeled as e_y, \mathrm{h}_{\mathrm{c}} is the characteristic of the signal, the electrical activity is termed as y_c. The diagnosis is observed in this case and it is symbolized as g_d, the patterns in the EEG signal are described as p^{\prime} the observation time is used to complete the detection of the signal and it is specified as \mathrm{O}_{\mathrm{e}}. The processing stage is used to examine the patterns and the characteristics of the EEG signal. Here, the computation process is used to deploy the electrical activity which is observed in this case.
The observation time is detected for this approach and provides reliable computing based on the diagnosis of epilepsy in the signal. The epilepsy identification is processed for the observation time where the characteristic is performed for the diagnosis of the EEG signal and it is represented as \left(\frac{{ }^{e^y} / p_{p^{\prime}+o_e}}{y_{\mathrm{c}} / \Pi_{\mathrm{h}_{\mathrm{c}}}\left(\mathrm{g}_{\mathrm{d}}\right)}\right). In this case, patterns and characteristics are associated with the observation time where the identification is performed for the electrical activity and it is symbolized as \left[\left(\mathrm{h}_{\mathrm{c}} * \frac{\mathrm{gd}}{e_{\mathrm{y}}}\right)+\left(\mathrm{p}^{\prime} * \mathrm{y}_{\mathrm{c}}\right)\right]. In this evaluation of epilepsy detection, the electrical activity is pragmatic for further computation. From this analysis phase, the extraction of the signal is performed in the below derivation.
\gamma=\left\{\frac{\left[\left(\beta * s_0\right) *\left(e_y * \frac{g_d}{a_i}\right)\right]}{\left(h_c+p^{\prime}\right)}\right\} * \sum_{h_c p_{p^{\prime}}}^{\left(s_{o^*} \frac{e_y}{a_i}\right)}\left[\left(o_e * e_y\right)+\left(\beta-y_c\right)\right]+\left(\begin{array}{c}s_0 * \frac{a_i}{h_c} / \sum_{e_y} \beta / o^{o_e}\end{array}\right) (2)
The above derivation states the extraction process of signal in which the amplitude is used to determine the early diagnosis of the epilepsy detection in EEG signal and it is symbolized as \gamma. The EEG signal is described as S_0, the amplitude detection in this equation and it is labeled as \mathrm{a}_{\mathrm{i}}. The process of computing the epilepsy is analyzed from the extraction phase where the characteristics and patterns are associated with the amplitude differentiation. This stage of evaluation of epilepsy is analyzed by extracting the necessary features from the input EEG signal. The amplitude differentiation is presented in Figure 2.
The input EEG signal characteristics are extracted as high/low for different patterns observed. However, based on intensity and frequency the p^{\prime} is different using y_c variations. Each p^{\prime} is detected between the start and end o_e such that \gamma is based on dissimilar h_c, In this case, the non-correlating y_c and h_c are identified as a_i between O_e. Therefore, the amplitude is detected between varying p^{\prime} other than high/low characteristics. This is performed to precisely categorize signals based on amplitude for diagnosis (Figure 2). The input signal is pragmatic with the characteristics and patterns where the observation time is analyzed in this case and it is represented as \left(s_0 * \frac{a_i}{h_c} / \sum_{e_y} \beta / \frac{o_e}{p^{\prime}}\right). The parameter of this equation is used to state the amplitude differentiation for the diagnosis of the diseases. This extraction is processed to observe the amplitude of the input image and based on this the computation time is observed for the further detection of the diseases. Post to this process the detection of amplitude is calculated in the further equation.
Figure 2. Amplitude differentiation
\mu=c_d * \frac{1}{s_n}+\left(h_c+p^{\prime}\right)+\sum_\beta^\gamma\left[\left(o_e+e_y\right)+\left(\frac{g_d}{y_c+m^{\prime}}\right)\right] * d_f+\left[\left(\frac{e_y}{h_c * p^{\prime}}\right) *\left(c_d+s_n\right) *\left(\frac{c_d+s_0}{\prod_{o_e}\left(a_i+e_y\right)}\right)\right] (3)
The detection of amplitude is processed in the above equation and it is termed as \mu. In this case, prediction is observed for the signal and it is symbolized as \mathrm{C}_{\mathrm{d}}, in this case, the number of signals which are been extracted from the input phase is labeled as S_n. The treatment is observed in this processing step and it is described as m^{\prime}, the differentiation is \mathrm{d}_{\mathrm{f}}. In this case, detection is based on the extraction of the signal where the amplitude differentiation is performed. The computation step here is pragmatic with the electrical activity in the brain to analyze the normal and abnormal detection these processes is carried with the diagnosis phase and it is equated as \left(\mathrm{o}_{\mathrm{e}}+\mathrm{e}_{\mathrm{y}}\right)+\left(\frac{\mathrm{g}_{\mathrm{d}}}{\mathrm{y}_{\mathrm{c}}+\mathrm{m}^{\prime}}\right).
Based on the detection of the normal the processing is stopped and provides the result. In another case, if it is abnormal then the identification of epilepsy is estimated with the electrical activity and provides the early stage of detection. Thus, the evaluation of identification of the diseases is calculated in the lesser observation time. Here, epilepsy detection is carried out for the characteristics and patterns from which the prediction is performed with the existing dataset and it is formulated as \left(\frac{\mathrm{e}_{\mathrm{y}}}{\mathrm{h}_{\mathrm{c}^* * \mathrm{p}^{\prime}}}\right) *\left(\mathrm{c}_{\mathrm{d}}+\mathrm{s}_{\mathrm{n}}\right). Deep learning is used in this work to differentiate between successive oscillations to correlate with the training data in detecting low signal outputs and it is discussed in the below section.
3.1 Deep learning for classification
Deep learning is associated with the artificial neural network and provides efficient computation based on the complex characteristics and patterns from the input signal. The processing step is used to deploy the prediction of the features in the EEG signal. Here, the treatment is used to examine epilepsy detection in the early stage of identification. The number of neurons is responsible for the ADM. Here, the low and high signal is associated with the diagnosis phase and the identification of the signal is used to deploy the low and high range of signal. The following equation is used to identify the signal which is low and high range of signal.
\alpha=\sum_{y_c}\left\{\left[\left(\frac{e_y}{d_f}+c_d / s_0\right)+\left(\beta * \frac{y_c+g_d}{o_e}\right)\right]\right\}+\prod_{c_d}^\mu\left[\left(h_c * p^{\prime}\right)+\left(s_0+\beta\right)\right] *\left(\gamma+\frac{c_d+e_y}{m^{\prime}}\right)-\left(d_f+a_i\right) (4)
The identification is derived in the above Eq. (4) where the epilepsy is detected based on the prediction and it is represented as \alpha. Here, the computation is based on the diagnosis of the observation time where the analysis is performed on the characteristics and patterns. The prediction is associated with the characteristics patterns and treatment and it is formulated as \left[\left(\mathrm{h}_{\mathrm{c}} * \mathrm{p}^{\prime}\right)+\left(\mathrm{s}_0+\beta\right)\right]. Here, the processing is analyzed is measured with the electrical activity and provides better detection of normal and abnormal phases of identification. The deep learning network is designed with two conditional analysis layers, one input and one output layer. The \mathrm{a}_{\mathrm{i}} is the input for which the \mathrm{d}_{\mathrm{f}} conditional validations are performed in the intermediate layers. In the consecutive second layer, m and \beta are the concurrent validations. The output is the sequence of different amplitudes observed based on variation and non-variations. The classification is performed for the \gamma variants to improve the iteration rates. The network is trained using the abnormal and normal sequences consecutively to improve the precision of detection. Besides, the optimization is induced through \mathrm{h}_{\mathrm{c}} classification; the procedure is defined as a partial derivative function to leverage the accuracy across different \rho.
The signal phase is detected with the low and high range of signal where the extraction phase is used to deploy the prediction. The processing step is used to identify the amplitude of the signal and provides the extraction of the necessary features for the treatment and it is represented as \left(\gamma+\frac{c_d+e_y}{m^{\prime}}\right). Thus, deep learning is used to identify the low and high range of signals for the forthcoming procedure. The neurons are interlinked with the identification of the EEG signal and from this examination are measured for the differentiation between the successive oscillation and it is equated below.
\rho=d_f+\left(l_0 * s_0\right)+\sum_{s_0}^{s_n}\left\lceil\left(y_c+\frac{h_c+p^{\prime}}{d_f}\right) *\left(c_d+\frac{a_i+d_f}{g_d / \mu+h_c}\right)\right] * \prod_{c_d}\left(\beta+y_c\right)-m^{\prime} (5)
The examination is performed in this derivation and it is equated as \rho, the detection is used to associate with the high and low signal and they are symbolized as \mathrm{s}_0 and \mathrm{l}_0. Here, the amplitude differentiation is examined between the successive oscillations. In this processing step, epilepsy detection is performed with amplitude differentiation and it is formulated as \left(\mathrm{y}_{\mathrm{c}}+\frac{\mathrm{h}_{\mathrm{c}}+\mathrm{p}^{\prime}}{\mathrm{d}_{\mathrm{f}}}\right). The diagnosis is associated with the analysis of the electrical activity and examines the better neuron computation. The DL is used to deploy the early diagnosis of epilepsy detection for the EEG signal. The deep learning process is illustrated in Figure 3.
In the learning process the \mathrm{p}^{\prime} is the first extraction for \varphi=\mathrm{d}_{\mathrm{f}} and \varphi \neq \mathrm{d}_{\mathrm{f}} differentiation. The differentiation process is required for m^{\prime} and \beta is dependent on validating the presence of \mathrm{s}_0 and \mathrm{l}_{\mathrm{o}} based on amplitude. The restoration is pursued for \varphi and \gamma for all s_n; the correlation process is required to validate if \mathrm{p}^{\prime} varies between successive \mathrm{O}_{\mathrm{e}}. The proposed identification is useful in training new amplitude variations. Considerably the training for both l_{\mathrm{o}} and \mathrm{s}_0 outputs are performed for new \mathrm{p}^{\prime} (detected) (Refer to Figure 3). The analysis is used to deploy the electrical activity and examine the predicting phase in this approach. Here, the treatment is observed to differentiate the amplitude the EEG signal. The computation step is used to examine the observation time where the low and high signals are differentiated. The prediction process is used to evaluate epilepsy detection in this DL where the characteristics and patterns are diagnosed for better neuron processing. From this evaluation process, the correlation factor is detected in the following equation.
Figure 3. Deep learning for \mathrm{a}_{\mathrm{i}} classification
\theta=\prod_{g_d}\left[\left(\left(h_c+p^{\prime}\right)+\frac{e_y}{s_0}\right)\right] *\left(c_d+\rho\right)-\left(y_c+\frac{a_i}{d_f}\right) (6)
The correlation factor is executed in the above equation and it is formulated as \theta. The diagnosis is performed for the characteristics and patterns and provides the signal processing to find the epilepsy and it is represented as \left(\left(h_c+p^{\prime}\right)+\frac{e_y}{s_0}\right). The early diagnosis of epilepsy provides the necessary treatment. In this computation process, the normal and abnormal are detected for the predicting phase. Here, the electrical activity is used to deploy the early diagnosis and it is associated with the DL in which neurons are processed for the ADM. The correlation factor is used to differentiate the normal and abnormal detection of EEG signals. The assessment layer is used to restore and un-restore oscillations are used in this approach and it is equated below
\left.\begin{array}{l}s_0=\sum_{l_0}^{h_0}\left(c_d+a_i\right) * \frac{o_e(0)}{g_d+\left(r_t+u_r\right)}-m^{\prime}+\beta \\ s_1=\sum_{l_0}^{h_0}\left(c_d+a_i\right) * \frac{o_e(1)}{g_d+\left(r_t+u_r\right)}-m^{\prime}+\beta \\ s_n=\sum_{l_0}^{h_0}\left(c_d+a_i\right) * \frac{o_e(n-1)}{g_d+\left(r_t+u_r\right)}-m^{\prime}+\beta\end{array}\right\} (7)
The assessment layer is responsible for the restoring and un-restoring oscillation and it is described as r_t and u_r. Here, the initial step of observation time is detected for the restoring and un-restoring process in which the normal and abnormal are identified. In this study, the precision is improved from this process where the numbers of neurons are associated with the analysis phase. In this category, the early diagnosis is detected with the training set where if there is any variation that occurs from high to low then the detection is not accurate. So the training phase is introduced for the correlation factor between the high and low factors. Post to this training data is performed in Deep learning and it is equated in the below equation as follows.
t_n=\left(\frac{\mu+\left(h_c+p^{\prime}\right)}{s_0 / \theta+e_y}\right) * \prod_{l_0}\left(\beta+s_n\right) *\left(a_i+d_f\right)-m^{\prime} (8)
From this training phase, the numbers of neurons are associated with the treatment improved for the diseases and it is labeled as \mathrm{t}_{\mathrm{n}}. The category of this EEG signal extraction is performed for the number of signals and from this amplitude differentiation is performed to analyze the diseases. The diagnosis is measured in this case for the examination of the successive oscillations where the restoring and un-restoring are observed. The calculation of restoring data is carried out by the predicting stage whereas, un-restoring is measured with the variation changes from high to medium and then low and vice versa. To avoid this variation, process the classification is derived in the following equation.
\partial=\left\{\begin{array}{c}\left(s_0+\gamma\right)+\left[\left(\frac{e_y+g_d}{\sum c_d+r_t}\right) * \beta+d_f\right], \in n^{\prime} \\ \left(\frac{\left(d_f+s_n\right)}{u_r * y_c}\right)-\left(m^{\prime}+\theta\right) * \mu, \in b^{\prime}\end{array}\right. (9)
The classification is observed in this case, for the normal and abnormal and they are represented as n'and \mathrm{b}^{\prime}. The classification is described as \partial; the observation that takes place for precision and accuracy enhancement.
In this category, the analysis is carried out for the amplitude differentiation where the epilepsy detection is processed. The computation relies on restoring the training in Eq. (8), so it is defined as normal. The second derivation is abnormal where the prediction is not observed where the correlation ranges in the high to medium value. The classification decision process is represented in Figure 4. In the decision process, the o value outputs are utilized for the types of \mathrm{d}_{\mathrm{f}} \forall \mathrm{r}_{\mathrm{t}} and \mu. If o \in n^{\prime} post the correlation process then \mathrm{d}_{\mathrm{f}} classifies either \mathrm{s}_0 or \mathrm{l}_{\mathrm{o}}. In the \partial occurring case the \mathrm{d}_{\mathrm{f}} is performed for \mu \forall \mathrm{s}_0 and \mathrm{l}_{\mathrm{o}} and this requires restoration checking. This differentiation for \mathrm{o} \in \mathrm{n}^{\prime} requires multiple classifications across \mathrm{s}_{\mathrm{n}} to predict further un-restoration cases. The \mathrm{u}_{\mathrm{r}} is identified for different inputs that are endorsed for \mathrm{d}_{\mathrm{f}}(\mu) under \alpha detection (Figure 4). From this approach, the differentiation of amplitude is measured for the restoring and unrestoring oscillations. This derivation is given below.
d_f=\left(s_0+\gamma\right) *\left(h_c+p^{\prime}\right)+\left(\frac{r_t+n^{\prime}}{\sum_{o_e}\left(e_y+c_d\right)}\right) * u_r\left(s_0\right) (10)
The observation is performed for the amplitude differentiation where the restoring and un-restoring are observed in this case. The computation process is examined for the normal signal where the restoring occurs in these layers in the high to low ranges observed. The un-restoring is detected when the value is low to medium and then high. To avoid the high impact in this process the differentiation for this restoring and un-restoring is measured for the oscillations in this proposal. From this identification is carried for the maximum correlation factor which is derived from Eq. (6).
\alpha=\sum_{e_y}\left(c_d+s_n\right) *\left(\frac{\gamma+\left(h_c+p^{\prime}\right)}{o_e+r_t}\right)+m^{\prime} * \beta (11)
The identification is carried out in this method and it is symbolized as \alpha,, in this category the prediction is performed for the extraction of the necessary information from the restoring data. Here, the analysis is examined for the restoration of the data in the assessment layer where the training is carried out on the mentioned observation time. In this process, extraction is carried out for the characteristics and patterns observed for the prediction of the restoring method and it is represented as \left(\frac{\gamma+\left(\mathrm{h}_{\mathrm{c}}+\mathrm{p}^{\prime}\right)}{\mathrm{o}_{\mathrm{e}}+\mathrm{r}_{\mathrm{t}}}\right). In this processing step, amplitude differentiation is detected for the successive oscillation in this process. The forthcoming method is to compute the observation time and reduce it in this work and it is equated in the below section.
\emptyset=\frac{1}{s_n}+\left(c_d+e_y\right) *\left(a_i+g_d\right)+\frac{\theta+\left(l_0+s_0\right)}{\prod_\gamma\left(\rho+r_t\right)} (12)
The computation is examined in the above equation and it is described as \varnothing, here the detection is performed for the restoring of the data in which the amplitude differentiation is carried out for the prediction. The correlation factor is used to examine the high and low range from which the restoration is extracted and it is equated as \frac{\theta+\left(l_0+s_0\right)}{\prod_\gamma\left(\rho+r_t\right)}. In this processing step diagnosis is carried out for the prediction method in which the extraction is observed for better precision in DL. Following this computation, the observation time is reduced and the verification phase is derived from the following formula.
\vartheta=\left(\alpha+e_y\right)-\left(y_c+c_d\right) *\left(\frac{\emptyset+h^{\prime}+t_n}{\sum_{g_d}\left(r_t * l_0\right)}\right)+\partial (13)
In the above equation, the identification of epilepsy is accomplished by the prediction and it is labeled as \vartheta, the phase is represented as \mathrm{n}^{\prime}. The detection process is carried out for the classification model in which the training is executed. The detection process using correlation output is presented in Figure 5.
In Figure 5 above the detection based on correlation outputs is performed. The proposed model validates d_f \forall r_t and \mu using \mathrm{s}_{\mathrm{o}} to \mathrm{s}_{\mathrm{n}} \forall \partial. This \partial ensures \mathrm{n}^{\prime} and \mathrm{b}^{\prime} using \mathrm{l}_{\mathrm{o}} and \mathrm{s}_0 in multiple t_n instances. If the \partial output is true then detection is pursued using n^{\prime} provided n^{\prime} \in h_c in either of l_o or s_0. If \alpha= low then the correlation for extracting r_t between different amplitudes is performed. This step differentiates r_t and u_r selection for new patterns in different o_e or trains the \gamma for new detection. The low signal outputs are verified until their next phase for their normal amplitude which is verified using DL under abnormal or normal classification. In this stage, the electrical activity is examined for the prediction of the restoring data in which the classification module is followed up. Thus, the characteristics and patterns are executed for the correlation factor. From this precision is improved from the analysis stage and it is formulated below.
\beta(\delta)=\sum_{s_0}^{s_n}\left[\left(h_c+p^{\prime}\right) *\left(\frac{c^{\prime}+n^{\prime}}{y_c}\right)\right]+(\alpha * \emptyset)-\tau (14)
The analysis is carried out for the precision in which the characteristics and patterns are executed for the betterment of the precision and it is \delta. The correlation time is observed in this method for the reliable computation of restoring and un-restoring data and it is expressed as \tau. In this stage, amplitude differentiation is performed for the detection of epilepsy in the early stage by using Deep learning. The scope of this work is satisfied by proposing DL and ADM methods. Thus, the correlation factor is maintained for the low signal extraction and finds the diseases. This proposed method exhibits maximum differentiations based on observation time from which the high precision unrestored oscillation improves the accuracy with less correlation time.
4.1 Experimental results
The results presented here are extracted from MATLAB experiments carried out using the dataset [33]. The dataset provides 256 EEG samples (approximately) observed from 22 human subjects for a maximum of 10s intervals. In this, 198 are identified as seizures and the remaining as non-seizure. The amplitude change range varies between -60 \mu \mathrm{~V} and +60 \mu \mathrm{~V} in the 1:1 aspect ratio at 50Hz. This configuration is analyzed in a physical machine with 2\times4GB random access memory and 1.8GHz processing speed. The proposed network is trained under 1200 iterations with 8 epochs. The training is performed with all the 256 inputs from which 150 are used for testing. The learning rate is set as 0.4 (minimum) and 1.0 (maximum) for different signal types. In particular, the normal signal type requires less iteration than the abnormal ones. The network is recurrently trained from the 800th iteration to the 1200th iteration irrespective of the samples. The ethical considerations are accounted for from the dataset used. This data utilization follows the standard and credential shared inheritance such that the shared rules are not violated. Therefore, for standard data utilization, the norms recommended in the dataset are followed to ensure no user credentials and EEG data is leaked. Besides, the EEG data is not exactly represented rather a partial part is extracted and used for assessment. The extraction is represented in the following experimental analysis. Based on the experimental outcomes, the results are presented in the Tables 1-4.
Table 1. Extraction
|
Signal Type |
Representation |
Extracted |
|
Normal |
||
|
Abnormal |
Table 2. Amplitude classification
|
Signal Type |
Low |
High |
|
Normal |
||
|
Abnormal |
Table 3. Differentiation
|
Type |
Differentiation\left(l_0\right) |
Differentiation\left(\boldsymbol{h}_o\right) |
|
Normal |
||
|
Abnormal |
Table 4. Detection
|
Type |
\boldsymbol{a}_i\left(h_c\right) |
Detected |
|
Normal |
||
|
Abnormal |
4.2 Comparative analysis
The comparative analysis discussed in this section comprises accuracy, precision, correlation factor, correlation time, and differentiation error. The above metrics are validated for a maximum observation time of 10s and maximum amplitude change of 50\mu \mathrm{V}. The above metrics are analyzed with the existing methods: mEEGNet [24], CNN-LSTM [32], and ConvLSTM [18]. In Table 5 and Table 6, the comparative analysis results are tabulated
The proposed AD model improves accuracy by 15.76%, precision by 15.11%, and correlation factor by 15.79%. This model reduces correlation time by 9.07% and differentiation error by 9.57%.
The proposed AD model improves accuracy by 15.55%, precision by 14.35%, and correlation factor by 16.08%. This model reduces correlation time by 9.55% and differentiation error by 10.78%.
Table 5. Comparative analysis results for observation time
|
Metrics |
mEEGNet |
CNN-LSTM |
ConvLSTM |
ADM-DL |
|
Accuracy |
0.693 |
0.782 |
0.869 |
0.9389 |
|
Precision |
0.728 |
0.791 |
0.867 |
0.9464 |
|
Correlation Factor |
0.752 |
0.835 |
0.934 |
0.9982 |
|
Correlation Time (ms) |
619.5 |
525.2 |
388.9 |
233.02 |
|
Differentiation Error |
0.105 |
0.084 |
0.075 |
0.0561 |
Table 6. Comparative analysis results for amplitude changes
|
Metrics |
mEEGNet |
CNN-LSTM |
ConvLSTM |
ADM-DL |
|
Accuracy |
0.702 |
0.782 |
0.852 |
0.9342 |
|
Precision |
0.741 |
0.802 |
0.875 |
0.9495 |
|
Correlation Factor |
0.754 |
0.841 |
0.917 |
0.9981 |
|
Correlation Time (ms) |
610.96 |
492.2 |
378.1 |
210.76 |
|
Differentiation Error |
0.109 |
0.086 |
0.076 |
0.0544 |
In this article, the ADM is introduced to improve the effective detection of epilepsy from EEG signals. The signal characteristics of high and low amplitudes were used to classify normal and abnormal inputs. The proposed model is aided by deep learning to verify the signal restorations despite amplitude changes. The amplitude restoration and non-restoration between the successive phases are identified for normal and abnormal signal detection. The high and low variations are induced for differentiation to achieve high-to-low or low-to-low retention of differential characteristics. The amplitude phases that do not align with the above differentiation are detected as seizures occurring between successive intervals. The identified changes are correlated with the training data to accurately detect seizures irrespective of the observation intervals. The learning model is trained using the observed restored and un-restored changes. The proposed AD model improves accuracy by 15.76%, precision by 15.11%, and correlation factor by 15.79%. This model reduces the correlation time by 9.07% and the differentiation error by 9.57% across various observation intervals. The extraction of signal characteristics requires the utilization of differential features across different observation times. This proposed model is less compatible with low observation intervals due to the lower precision achieved. To address this issue, the implementation of a regressive probability distribution function is required and is being considered for future work.
[1] Supriya, S., Siuly, S., Wang, H., Zhang, Y. (2021). Epilepsy detection from EEG using complex network techniques: A review. IEEE Reviews in Biomedical Engineering, 16: 292-306. https://doi.org/10.1109/RBME.2021.3055956
[2] Tasci, I., Tasci, B., Barua, P.D., Dogan, S., Tuncer, T., Palmer, E.E., Fujita, H., Acharya, U.R. (2023). Epilepsy detection in 121 patient populations using hypercube pattern from EEG signals. Information Fusion, 96: 252-268. https://doi.org/10.1016/j.inffus.2023.03.022
[3] Wong, S., Simmons, A., Rivera-Villicana, J., Barnett, S., Sivathamboo, S., Perucca, P., Kwan, P., Kuhlmann, L., Vasa, R., O'Brien, T.J. (2023). EEG based automated seizure detection-A survey of medical professionals. Epilepsy & Behavior, 149: 109518. https://doi.org/10.1016/j.yebeh.2023.109518
[4] Rukhsar, S. (2022). Discrimination of multi-class EEG signal in phase space of variability for epileptic seizure detection using error correcting output code (ECOC). International Journal of Information Technology, 14(2): 965-977. https://doi.org/10.1007/s41870-018-0224-y
[5] Nielsen, J.M., Zibrandtsen, I.C., Masulli, P., Sørensen, T.L., Andersen, T.S., Kjær, T.W. (2022). Towards a wearable multi-modal seizure detection system in epilepsy: A pilot study. Clinical Neurophysiology, 136: 40-48. https://doi.org/10.1016/j.clinph.2022.01.005
[6] Pan, Y., Dong, F., Wu, J., Xu, Y. (2023). Downsampling of EEG signals for deep learning-Based epilepsy detection. IEEE Sensors Letters, 7(12): 1-4. https://doi.org/10.1109/LSENS.2023.3332392
[7] Ueda, T., Iimura, Y., Mitsuhashi, T., Suzuki, H., Miao, Y., Nishioka, K., Tamrakar, S., Matsui, R., Tanaka, T., Otsubo, H., Sugano, H., Kondo, A. (2023). Chronological changes in phase-amplitude coupling during epileptic seizures in temporal lobe epilepsy. Clinical Neurophysiology, 148: 44-51. https://doi.org/10.1016/j.clinph.2023.01.014
[8] Lih, O.S., Jahmunah, V., Palmer, E.E., Barua, P.D., Dogan, S., Tuncer, T., García, S., Molinari, F., Acharya, U.R. (2023). EpilepsyNet: Novel automated detection of epilepsy using transformer model with EEG signals from 121 patient population. Computers in Biology and Medicine, 164: 107312. https://doi.org/10.1016/j.compbiomed.2023.107312
[9] Tuncer, E., Bolat, E.D. (2022). Channel based epilepsy seizure type detection from electroencephalography (EEG) signals with machine learning techniques. Biocybernetics and Biomedical Engineering, 42(2): 575-595. https://doi.org/10.1016/j.bbe.2022.04.004
[10] El-Gindy, S.A.E., Hamad, A., El-Shafai, W., Khalaf, A.A., El-Dolil, S.M., Taha, T.E., El-Fishawy, A.S., Alotaiby, T.N., Alshebeili, S.A., El-Samie, F.E.A. (2021). Efficient communication and EEG signal classification in wavelet domain for epilepsy patients. Journal of Ambient Intelligence and Humanized Computing, 12: 9193-9208. https://doi.org/10.1007/s12652-020-02624-5
[11] Tamrakar, S., Iimura, Y., Suzuki, H., Mitsuhashi, T., Ueda, T., Nishioka, K., Karagiozov, K., Nakajima, M., Miao, Y., Tanaka, T., Sugano, H. (2022). Higher phase-amplitude coupling between ripple and slow oscillations indicates the distribution of epileptogenicity in temporal lobe epilepsy with hippocampal sclerosis. Seizure, 100: 1-7. https://doi.org/10.1016/j.seizure.2022.05.024
[12] Kumar, A., Lyzhko, E., Hamid, L., Srivastav, A., Stephani, U., Japaridze, N. (2021). Differentiating ictal/subclinical spikes and waves in childhood absence epilepsy by spectral and network analyses: A pilot study. Clinical Neurophysiology, 132(9): 2222-2231. https://doi.org/10.1016/j.clinph.2021.06.011
[13] Fukumori, K., Yoshida, N., Sugano, H., Nakajima, M., Tanaka, T. (2021). Epileptic spike detection using neural networks with linear-phase convolutions. IEEE Journal of Biomedical and Health Informatics, 26(3): 1045-1056. https://doi.org/10.1109/JBHI.2021.3102247
[14] Monsoor, T., Zhang, Y., Daida, A., Oana, S., Lu, Q., Hussain, S.A., Fallah, A., Sankar, R., Staba, R.J., Speier, W., Roychowdhury, V., Nariai, H. (2023). Optimizing detection and deep learning-based classification of pathological high-frequency oscillations in epilepsy. Clinical Neurophysiology, 154: 129-140. https://doi.org/10.1016/j.clinph.2023.07.012
[15] Yu, J., Qiu, K., Wang, P., Su, C., Fan, Y., Cao, Y. (2023). Perturbing BEAMs: EEG adversarial attack to deep learning models for epilepsy diagnosing. BMC Medical Informatics and Decision Making, 23(1): 115. https://doi.org/10.1186/s12911-023-02212-5
[16] Srinivasan, S., Dayalane, S., Mathivanan, S.K., Rajadurai, H., Jayagopal, P., Dalu, G.T. (2023). Detection and classification of adult epilepsy using hybrid deep learning approach. Scientific Reports, 13(1): 17574. https://doi.org/10.1038/s41598-023-44763-7
[17] Jeon, Y., Chung, Y.G., Joo, T., Kim, H., Hwang, H., Kim, K.J. (2022). Deep learning-based detection of epileptiform discharges for self-limited epilepsy with centrotemporal spikes. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 30: 2939-2949. https://doi.org/10.1109/TNSRE.2022.3215526
[18] Tawhid, M.N.A., Siuly, S., Li, T. (2022). A convolutional long short-term memory-based neural network for epilepsy detection from EEG. IEEE Transactions on Instrumentation and Measurement, 71: 1-11. https://doi.org/10.1109/TIM.2022.3217515
[19] De Brabandere, A., Chatzichristos, C., Van Paesschen, W., De Vos, M., Davis, J. (2023). Detecting epileptic seizures using hand-crafted and automatically constructed EEG features. IEEE Transactions on Biomedical Engineering, 71(1): 318-325. https://doi.org/10.1109/TBME.2023.3299821
[20] Pandya, V., Shukla, U.P., Joshi, A.M. (2023). Novel features extraction from EEG signals for epilepsy detection using machine learning model. IEEE Sensors Letters, 7(10): 1-4. https://doi.org/10.1109/LSENS.2023.3309254
[21] Lee, S.Y., Hung, Y.W., Chang, Y.T., Lin, C.C., Shieh, G.S. (2021). RISC-V CNN coprocessor for real-time epilepsy detection in wearable application. IEEE Transactions on Biomedical Circuits and Systems, 15(4): 679-691. https://doi.org/10.1109/TBCAS.2021.3092744
[22] Wang, Q., Kong, W., Zhong, J., Shan, Z., Wang, J., Li, X., Hu, B. (2023). A hybrid SVM and kernel function-based sparse representation classification for automated epilepsy detection in EEG signals. Neurocomputing, 562: 126874. https://doi.org/10.1016/j.neucom.2023.126874
[23] Shen, M., Wen, P., Song, B., Li, Y. (2022). An EEG based real-time epilepsy seizure detection approach using discrete wavelet transform and machine learning methods. Biomedical Signal Processing and Control, 77: 103820. https://doi.org/10.1016/j.bspc.2022.103820
[24] Shoji, T., Yoshida, N., Tanaka, T. (2021). Automated detection of abnormalities from an EEG recording of epilepsy patients with a compact convolutional neural network. Biomedical Signal Processing and Control, 70: 103013. https://doi.org/10.1016/j.bspc.2021.103013
[25] Chakraborty, M., Mitra, D. (2021). Epilepsy seizure detection using kurtosis based VMD’s parameters selection and bandwidth features. Biomedical Signal Processing and Control, 64: 102255. https://doi.org/10.1016/j.bspc.2020.102255
[26] Cao, J., Grajcar, K., Shan, X., Zhao, Y., Zou, J., Chen, L., Li, Z., Grunewald, R., Zis, P., De Marco, M., Unwin, Z., Blackburn, D., Sarrigiannis, P.G. (2021). Using interictal seizure-free EEG data to recognize patients with epilepsy based on machine learning of brain functional connectivity. Biomedical Signal Processing and Control, 67: 102554. https://doi.org/10.1016/j.bspc.2021.102554
[27] Lebal, A., Moussaoui, A., Rezgui, A. (2023). Epilepsy-Net: attention-based 1D-inception network model for epilepsy detection using one-channel and multi-channel EEG signals. Multimedia Tools and Applications, 82(11): 17391-17413. https://doi.org/10.1007/s11042-022-13947-0
[28] Majzoub, S., Fahmy, A., Sibai, F., Diab, M., Mahmoud, S. (2023). Epilepsy detection with multi-channel EEG signals utilizing Alexnet. Circuits, Systems, and Signal Processing, 42(11): 6780-6797. https://doi.org/10.1007/s00034-023-02423-1
[29] Goel, S., Agrawal, R., Bharti, R.K. (2024). Automated detection of epileptic EEG signals using recurrence plots-based feature extraction with transfer learning. Soft Computing, 28(3): 2367-2383. https://doi.org/10.1007/s00500-023-08386-4
[30] Rajinikanth, V., Kadry, S., Taniar, D., Kamalanand, K., Elaziz, M.A., Thanaraj, K.P. (2023). Detecting epilepsy in EEG signals using synchro-extracting-transform (SET) supported classification technique. Journal of Ambient Intelligence and Humanized Computing, 14(8): 10123-10141. https://doi.org/10.1007/s12652-021-03676-x
[31] Sunaryono, D., Sarno, R., Siswantoro, J. (2022). Gradient boosting machines fusion for automatic epilepsy detection from EEG signals based on wavelet features. Journal of King Saud University-Computer and Information Sciences, 34(10): 9591-9607. https://doi.org/10.1016/j.jksuci.2021.11.015
[32] Pandey, S.K., Janghel, R.R., Mishra, P.K., Ahirwal, M.K. (2023). Automated epilepsy seizure detection from EEG signal based on hybrid CNN and LSTM model. Signal, Image and Video Processing, 17(4): 1113-1122. https://doi.org/10.1007/s11760-022-02318-9
[33] EEG Seizure Analysis Dataset. https://www.kaggle.com/datasets/adibadea/chbmitseizuredataset.
