A significant framework for epilepsy and seizure clusters detection using MSRS-LSTM and FID-DO

Abstract

Epilepsy is a neurological condition that causes unprovoked and recurring seizures. Electroencephalography (EEG) is essential for diagnosing epilepsy, as EEG records the brain's electrical activity. However, the prevailing works are not focused on detecting the multiple seizure occurrences in one person and the Seizure Clusters (SC). So, a significant epilepsy detection framework using Meta-step Rootsig Long Short-Term Memory (MsRs-LSTM) and Fractional integro-differential Duffing Oscillator (Fid-DO) is proposed in this framework. The EEG signals are initially preprocessed and decomposed using Extrema mirror Empirical mode Decomposition (EED). Then, the signals’ peaks are detected, sub-bands are identified, and features are extracted from them. Similarly, from the preprocessed signal, Fid-DO detects weak periodicity, and the features are extracted from the signals being detected. Afterwards, the extracted features’ dimensionality is reduced and given to MsRs-LSTM for epilepsy classification. The evaluation outcomes stated the robustness of the proposed framework in epilepsy detection with 97.56% classification accuracy.

Keywords

Epilepsy multiple seizures detection seizure clusters EEG signal MsRs-LSTM Fid-DO principal component analysis and damping error-term notch filter (DE-NF)

1. Introduction

Epilepsy is a neurological disorder that is exemplified by recurrent seizures.¹ These seizures are due to the brain's abnormal electrical activity.² Epilepsy can affect people of all ages, and its seriousness can vary extensively from person to person.³ Epilepsy can have numerous causes, including brain tumours, genetic factors, developmental disorders, brain injury, infections, and prenatal injuries.⁴ In many cases, the cause may not be identified, and it is named idiopathic epilepsy.⁵ There are 2 sorts of seizures namely, partial seizure, which affects the brain's particular area, and generalized seizure, which affects the entire brain.⁶ The symptoms of epilepsy differ based on the seizure types, and some common symptoms are temporary confusion, loss of consciousness, sensory changes, and automatic behaviours.⁷ Epilepsy is diagnosed by neurological examination through EEG.

Nevertheless, epilepsy cannot be cured in most cases.⁸ Several people effectively manage their condition with treatment plans such as getting enough sleep, managing stress, avoiding triggers, and taking medications as prescribed to minimize the risk of seizures.^9,10 A neurological condition that affects more than 50 million individuals worldwide, epilepsy is characterized by recurring seizures brought on by aberrant brain activity. Treatment options include anti-epileptic medications, surgery, ketogenic diets, and therapies; diagnosis is made using EEG, imaging, and medical history. Wearable technology, genetic research, and artificial intelligence advancements enhance diagnosis and treatment.

Early diagnosis and appropriate treatment are crucial for optimizing and minimizing the impact of seizures.¹¹ Several Machine Learning (ML) and Deep Learning (DL) methodologies like Convolutional Neural Network (CNN), Support Vector Machine (SVM), Artificial Neural Network (ANN), Recurrent Neural Network (RNN), and Random Forests (RF) are developed and employed for an early diagnosis of epilepsy.^12,13 However, the existing works didn’t effectively detect the epileptic signals from the weak periodicity,¹⁴ and all sub-bands were also not analyzed,¹⁵ which is insufficient for epileptic detection. So, an effective epilepsy detection model using MsRs-LSTM and Fid-DO is proposed in this framework.

1.1. Problem statement

Some major challenges of the existing works are as follows,

None of the existing works focused on detecting multiple seizures occurring in one person and seizure clusters.

The existing work utilized the raw EEG signal for detecting epilepsy, which resulted in inaccurate results.¹⁶

The existing work concentrated only on the alpha and delta sub-bands for epilepsy detection but failed to focus on the beta and gamma sub-bands.

The existing works didn’t effectively detect the epileptic signals from the weak periodicity, which is insufficient for epileptic detection.¹⁷

1.2. Objectives

The contributions of the proposed framework in epilepsy detection are described below,

The proposed framework effectively detected multiple seizures and seizure clusters by using MsRs-LSTM.

The EEG signals are preprocessed under normalization, noise removal by DE-NF, and contrast enhancement processes.

This framework detects the alpha, beta, gamma, and delta sub-bands for epilepsy detection.

Here, the weak periodicity is detected by using Fid-DO.

The paper's remaining organization is as follows: section 2 presents the literature survey, section 3 defines the proposed model, section 4 demonstrates the results, and section 5 winds up the paper.

2. Literature survey

Mitsis et al.¹⁸ explained the epilepsy patients’ Functional Brain Networks (FBNs), exposing pronounced multiscale periodicities. The hypothesis was assessed by utilizing long-duration scalp EEG data (21–94 h.) in 9 patients with epilepsy regarding which FBNs were developed. The outcomes showcased that these networks regularly differ over time, thus exhibiting several peaks at intervals ranging from 1 to 24 h. The seizure onset effects on the FBN features were established to be significantly smaller in magnitude in contrast to the alteration that happened owing to these inherent periodic cycles. In many of these subjects, a moderately lower number of seizures was documented.

Durga Praveen Devi (2024) uses machine learning approaches to increase the precision of MRI data for brain tumour diagnosis. The method uses a multilayer perceptron for classification, local phase quantization for texture extraction, graphing for structural analysis, and K-nearest neighbours preprocessing. On a test set of 1104 photos, the model's accuracy was 87%, with precision and recall scores of 0.86 and 0.93, respectively. This demonstrates how integrated machine learning may be used to diagnose illnesses.¹⁹

Akbarian and Erfanian²⁰ described a technique for seizure identification utilizing graph theory, effectual connectivity, and Multi-Level Modular Networks (MLMN). Regarding the amalgamation of numerous EBC classification outcomes at varied frequencies, the MLMN was implemented. The classification outcomes of 3 varied EBCs were combined at a particular frequency by utilizing another methodology termed Modular Effective Neural Network (MENN). Results revealed that the MLMN had the highest mean accuracy. In contrast to the other studies, which utilized the MIT-CHB database, the applied methodology offered higher accuracy. Even though the applied method had numerous benefits, the MLMN method indicated more time during the training.

Supriya et al.²¹ individualized the description of SC regarding temporal clustering evaluation. For the detection of automated individualized seizure clusters, an algorithm has been developed that amalgamated cumulative sum change-point evaluation with bootstrapping as well as aberration identification, thus providing a model for personalized seizure cluster detection at the clinical significance's user-specified levels. More seizures were identified as clusters by using the applied algorithm on the basis of routine definitions, which had the better probability. The limitation was that there were errors in identifying SC that might not satisfy a pre-specified threshold (poor sensitivity).

Patient care has improved due to the adoption of IoT technology in healthcare systems. Naresh Kumar Reddy Panga (2022) investigates how an Internet of Things-based health monitoring system processes ECG signals using the Discrete Wavelet Transform (DWT). The DWT's time-frequency localization capabilities are essential for effective non-stationary signal analysis. Preprocessing, feature extraction, signal capture, and real-time IoT-based transmission to cloud servers are all included in the system design. According to performance measurements, data reduction and signal clarity have improved.²²

Zhan and Hu²³ explained the feature extraction process for the epileptic seizures’ automated detection utilizing a complex network structure. The study's objective was to implement a graph theory-centric innovative model together with a complex network system that was effectual for the EEG signals’ automated classification to identify epilepsy. The evaluation showed that for the Bern-Barcelona database as well as the Bonn University database, a better accuracy level has been obtained. The experiential outcomes revealed that the applied model was more effective in distinguishing epileptic seizure signals as of varied EEG signals. Since the graph equation approach was included, there will be a possible information loss for the reason that the criteria forming the valuable nodes could be ignored.

Using the Hadoop Distributed File System (HDFS) Hadoop architecture and Improved You Only Look Once version four (YOLO v4), a novel technique for identifying small blood cells in acute lymphoblastic leukaemia (ALL) has been created. The enhanced YOLOv4 method resolves the ALLIDB1 dataset's class imbalance problems, increasing the accuracy of healthy and blast cell identification and recognition.²⁴

Sukriti et al.²⁵ described the epilepsy identification methodology utilizing a Multiview clustering algorithm together with deep features. This study applied a classification methodology regarding unsupervised Multiview clustering outcomes to improve the effect of automatic detection. The experiential outcomes demonstrated that, in contrast to the existing EEG identification, the applied system was highly intelligent, automated, and effective in terms of ML techniques. Analyzing the brain signals posed challenges because of their high volume of data.

Najafi et al.²⁶ explained the localization methodology for epileptic seizure onset zones regarding time frequency and clustering exploration. A threshold methodology was utilized to identify Events of Interest (EoIs). Next, for the purpose of acquiring Channels of Interest (CoIs) by computing the average power of EoIs on every single channel, a time-frequency evaluation model was employed. The localization model's better performance was displayed by analogizing its sensitivity and specificity with the other prevailing methodologies. Relying solely on seizure semiology (observed clinical features) may not accurately determine the site of seizure onset.

Aslam et al.²⁷ described the automated identification of epileptic seizures utilizing multiscale and refined composite Multiscale Dispersion Entropy (MDE). MDE and Refined Composite MDE (RCMDE), the 2 entropy features being established recently, were utilized to detect seizures. The experimental outcomes illustrated that both the MDE and RCMDE were regarded as promising feature extraction methodologies since they successfully quantify the EEG signals’ complications and obtain better classification accuracies whilst c = 5 and m = 3 for both MDE and RCMDE models. Accurate seizure detection required prolonged monitoring, introducing complexities due to multiple channels and storage.

3. Proposed methodology for detecting epilepsy using msrs-lstm and fid-do

In this framework, the epilepsy disease is detected using MsRs-LSTM and Fid-DO. EEG signals and other sequential data are processed by the recurrent Long Short-Term Memory (LSTM) neural network. It uses sophisticated processes to increase performance, decrease overfitting, and boost convergence and learning efficiency, including meta-step gradient pruning and rootsig activation. The model categorizes EEG data using sequential patterns into normal and epileptic states. Fid-DO is a mathematical model for detecting weak periodicities in EEG signals that is based on the Duffing Oscillator and improved with fractional integro-differential equations. It assists in detecting faint periodic patterns in EEG data frequently linked to seizures that linear approaches could overlook. Fractional derivatives improve the model's sensitivity to long-term dependencies and minute changes. Here, multiple seizures and seizure clusters are also detected. The structure of the proposed framework is depicted in Figure 1. The proposed approach employs cutting-edge methods to identify seizure and epilepsy clusters in EEG recordings. Signals are preprocessed using normalization, noise reduction, and contrast enhancement. They are then broken down into smaller parts and peaks and frequency sub-bands are identified. PCA is used to extract and minimize important characteristics, while a Fractional Integro-Differential Duffing Oscillator (Fid-DO) is used to identify weak periodicities. Features of the Meta-step Rootsig Long Short-Term Memory (MsRs-LSTM) model are categorized as either normal or epileptic.

Figure 1.

Architecture of the proposed framework.

The epilepsy detection process by the proposed framework is summarized as follows. Initially, the EEG signals are preprocessed and decomposed using EED. Then, the peaks and sub-bands are detected, followed by extracting features from them. Similarly, the Fid-DO detects weak periodicity from the preprocessed signal, and features are extracted from it. Next, the dimensions of the extracted features are reduced by PCA and given to MsRs-LSTM for epilepsy detection. Feature dimension suppression improves computational efficiency, model performance, and robustness by reducing the number of features in a dataset while maintaining important information. It reduces noise and redundancy, streamlines models, lessens the impact of the curse of dimensionality and facilitates visualization. By concentrating on key seizure-related characteristics, PCA is utilized in the suggested framework to optimize the EEG data for effective and precise epilepsy identification.

3.1. Preprocessing

Initially, the EEG signals are collected from the dataset and the data are preprocessed to augment the epilepsy detection's efficacy. EEG signals are scaled to a standard range using normalization, guaranteeing signal consistency and comparability. For example, it modifies the EEG signal's amplitudes to fall within a specified range (0 to 1 or −1 to 1). By doing this, biases brought on by differences in signal amplitudes are removed, and algorithms operate more consistently across datasets. The utilized e number of EEG signals $(E)$ is represented as,

\begin{aligned} E = {E_{1}, E_{2}, \dots \dots, E_{e}} \end{aligned}

(1)

The evaluation of the noise reduction stage guarantees that undesirable artefacts are eliminated while maintaining crucial characteristics, improving the quality of the EEG data. As a result, there are fewer false positives or negatives, enhanced detection accuracy, and dependable feature extraction. A key component of the suggested epilepsy detection architecture, noise reduction, is validated by SSIM, MAE, and PSNR metrics. The EEG signals are then preprocessed under normalization, noise removal, and contrast enhancement processes described below.²⁸

(i). Normalization: Normalization $(ε)$ of the EEG is considered a process of adjusting the amplitude of a signal into a specific range or scale, and it is represented as,

\begin{aligned} E \overset{ε}{⟶} E_{n o r} \end{aligned}

(2)

Here, $E_{n o r}$ denotes the normalized EEG signal.

(ii). Noise removal: Then the noises that occur $E_{n o r}$ are removed by DE-NF. The Notch Filter (NF) effectively removes unwanted frequencies or noises from a signal. NF reduces noise by attenuating specific frequencies while leaving others unaffected.²⁹ However, the occurrence of passband ripple causes fluctuations in the amplitude of frequencies near the notch frequency. This ripple can degrade the signal quality. So, a Damping error is used for compensation. The processes of DE-NF are as follows. Initially, the notch frequency $(Φ)$ , which is considered as the unwanted frequency component is determined by,

\begin{aligned} Φ = ℓ * \frac{c f}{s f} \end{aligned}

(3)

Here, $ℓ$ the Nyquist frequency $c f a n d s f$ represents the centre and sample frequencies, respectively. Then, the Damping error ratio $(ς)$ is calculated to compensate for the pass band ripple that occurs in the filter, and it is described as,³⁰

\begin{aligned} ς = - k τ + \frac{1}{2} Ψ (\frac{1}{ρ}) \sin 2 (ς - ς_{ρ}) \end{aligned}

(4)

Where, $τ$ denotes the precession angular rate, $Ψ$ represents the damping error rate, $ρ$ depicts the damping force rate, and k denotes the damping factor. Next, the coefficients $(Ξ)$ for the analog prototype are estimated using the quality factor $(Q)$ , and it is depicted as,

\begin{aligned} Ξ = \frac{\sin (Φ)}{2 Q} \end{aligned}

(5)

Then, a bilinear transform $(B T)$ is applied as,

\begin{aligned} B T = 2 * \tan (Ξ) \end{aligned}

(6)

Afterwards, the notch filter transfer function $(ℏ)$ is obtained with the input signal by convolving the signal with the filter coefficients, and it is determined as,

\begin{aligned} ℏ = E_{n o r} \cdot \frac{(χ_{0} + χ_{1} * z^{- 1} + χ_{2} * z^{- 2})}{1 + γ_{1} * z^{- 1} + γ_{2} * z^{- 2}} \end{aligned}

(7)

Where, z denotes the complex variable in z-domain, $χ_{0}, χ_{1}, χ_{2}$ are the coefficients for the input signal, and $γ_{1}, γ_{2}$ are the coefficients for the output signal. Thus, this notch filter effectively attenuates the notch frequency and allows the other frequencies to pass through the filter. Thus, the noise-removed signal is denoted as $E_{n o i}$ .

(iii). Contrast enhancement: The contrast of the noise-removed signal $(E_{n o i})$ is enhanced here to improve the visualization and highlight the features, and it is denoted as $E_{c o n}$ .

3.2. Signal decomposition

After that, the contrast-enhanced signal $(E_{c o n})$ is decomposed by EED. The Empirical-mode Decomposition (ED) retains the energy conservation property. It ensures that the sum of energies of the decomposed Intrinsic Mode Function (IMF) equals the energy of the original signal. This property makes EMD suitable for applications where signal energy preservation is important. However, EMD is sensitive to endpoint effects, particularly at the boundaries of the signal. The decomposition near the endpoints of the signal may be less accurate, leading to the generation of spurious artefacts. So to overcome this issue, an extrema mirror extension is used. Techniques such as the Fourier Transform may be used to analyze EEG data in the frequency domain, although they are usually represented in the time domain. The temporal and spectral features of the signal are captured by peak detection analysis after the signal is divided into its oscillatory components using Empirical Mode Decomposition (EMD). First, the input signal $(E_{c o n})$ is represented with time-domain $(T)$ as $E_{c o n} (T)$ , which is depicted as,

\begin{aligned} E_{c o n} (T) = \sum (℘_{i m} (T) + ψ_{i m} (T)) \end{aligned}

(8)

Here, $℘_{i m} (T) a n d ψ_{i m} (T)$ denote the IMF and residue, respectively; then, discover the local minima and maxima $E_{c o n} (T)$ . Afterwards, interpolate between the local minima and maxima for obtaining the upper $U E (T)$ and lower $L E (T)$ envelopes $E_{c o n} (T)$ , and it is depicted as,

\begin{aligned} I (T) = \frac{U E (T) + L E (T)}{2} \end{aligned}

(9)

Where, $I (T)$ denotes the interpolation. To enhance the decomposition accuracy of the endpoints, extrema mirror extension is utilized here, and it is depicted as,

\begin{aligned} U E (T) & = {U E (1), U E (2), \dots \dots ., U E (P)}, P \in p \end{aligned}

(10)

\begin{aligned} L E (T) & = {L E (1), L E (2), \dots \dots ., L E (P)}, P \in p \end{aligned}

(11)

Here, p denotes the total P number of frequency components. Then, from the envelopes, get the middle value $(M^{'} (T))$ and decrease it from the original signal to attain the first component $(Δ (T))$ , which is represented as,

\begin{aligned} Δ (T) = E_{c o n} (T) - M^{'} (T) \end{aligned}

(12)

The above process results in the first IMF $(ƛ_{1} (T))$ , representing the high-frequency oscillations $E_{c o n} (T)$ . Then, evaluate the first residue $(R_{1} (T))$ , which represents the low-frequency component $E_{c o n} (T)$ that has to be removed, and it is described as,

\begin{aligned} R_{1} (T) = E_{c o n} (T) - ƛ_{1} (T) \end{aligned}

(13)

Finally, the decomposed signal $(E_{d e} (T))$ is obtained by the sum of all IMFs and the last residue, and it is depicted as,

\begin{aligned} E_{d e} (T) = \sum_{p = 1}^{P} ƛ_{p} (T) + R_{P} (T) \end{aligned}

(14)

Then, the decomposed EEG signal is further processed for detecting the peaks to determine the sub-bands of the EEG signal.

3.3. Peak detection

Here, the peaks of the decomposed signals $(E_{d e} (T))$ are detected by estimating separate boxes representing the high amplitudes and every single peak. The peak detected signal is denoted as $E_{p e a k}$ .

3.4. Sub-band detection

From the peak detected signal $(E_{p e a k})$ , the sub-bands like alpha, beta, gamma, and delta are detected in every single peak indicated box, and it is represented as $E_{s u b}$ . The $E_{s u b}$ is determined by the frequency of $E_{p e a k}$ , which is given as,

\begin{aligned} E_{s u b} = {\begin{cases} 8 - 12 H z, & a l p h a \\ 13 - 29 H z, & b e t a \\ 30 - 60 H z, & g a m m a \\ 2 - 4 H z, & d e l t a \end{cases} \end{aligned}

(15)

3.5. Weak periodicity detection

Similarly, $E_{c o n}$ the weak periodicity is detected by Fid-DO. The Duffing Oscillator (DO) captures non-linear behaviour, making DO applicable to systems where linear models fail to represent the dynamics accurately. But DO is sensitive to initial conditions. Small disparities in initial conditions can lead to significantly diverse trajectories, making predictions challenging, especially for long-term behaviour. So, a fractional integro-differential equation with a nonlocal initial condition is used. At first, the initial condition of the Fid-DO is determined by the fractional integro-differential equation, and it is depicted as,

\begin{aligned} δ = {\overset{⌢}{ℜ}}^{α} R (r) + \int_{0}^{r} K (r, w) R (w) d w \end{aligned}

(16)

Here, $δ$ denotes the forcing function, $K (r, w)$ is the kernel function that depends on both current time $(r)$ and past time $(w)$ , $R (r) a n d R (w)$ represents the unknown function of $r a n d w$ , respectively, and ${\overset{⌢}{ℜ}}^{α}$ denotes the fractional derivative operator. The initial velocity, acceleration, and displacement are based on $δ$ . Then, the Fid-DO equation is given as,

\begin{aligned} \ddot{x} + λ \dot{x} + ϕ x + υ x^{3} = \nabla \cos (θ \hat{t}) \end{aligned}

(17)

Here, x denotes the displacement of the oscillator, $\dot{x} a n d \ddot{x}$ represents the velocity and acceleration of x, respectively, $λ, ϕ, υ$ denote the damping coefficient, linear stiffness coefficient, and nonlinearity parameter, respectively, $\cos (θ t)$ denotes the driving force, $\nabla$ is the amplitude of the driving force, $θ$ is the angular velocity, and $\hat{t}$ is time. Then, the signal is reconstructed by the inputted signal $E_{c o n}$ s, the driving force, and it is represented as,

\begin{aligned} \ddot{x} + λ \dot{x} + ϕ x + υ x^{3} = \nabla E_{c o n} \end{aligned}

(18)

Next, the numerical solution for Fid-DO is obtained by discretizing this system using a numerical integration method to get the displaced x. Then, the frequency spectrum $X (F)$ of the displaced x is obtained from Fid-DO, and it is indicated as,

\begin{aligned} X (F) = x \end{aligned}

(19)

Afterwards, weak periodicity is detected by analyzing the frequency peaks of the signals. The peak detection $(\tilde{ρ})$ is given as,

\begin{aligned} \tilde{ρ} \to (X (F))_{p e a k} \end{aligned}

(20)

The amplitudes of the detected peaks are analyzed to assess the strength of periodicity. The low amplitude indicates the weak periodicity $(κ)$ of the signal, which is represented as,

\begin{aligned} κ = \tilde{ρ} \in s \tilde{ρ}, i f ℘_{a m p} < ℘_{s i g} \end{aligned}

(21)

Here,

s \tilde{p}

denotes the significant peaks

℘_{a m p} a n d ℘_{s i g}

are the amplitudes of the detected peak and significant amplitude, respectively. Thus, the weak periodicity

(E_{w e a k})

is detected and the pseudocode for Fid-DO is presented below.

Hereafter, the features of the weak periodicity signals and sub-bands are extracted to train the epilepsy classification model.

3.6. Feature extraction

Feature extraction is extracting important features from EEG signals, such energy, mean and standard deviation, fracture dimensions, and Lyapunov exponents, to precisely determine the underlying patterns and differentiate between epileptic and normal states. From $E_{s u b}$ and $E_{w e a k}$ , the features like energy, standard deviation, mean, absolute value, average power, the ratio of absolute mean values, relative energy, line length, energy entropy, relative amplitude, coefficient of variation entropies, correlation dimension, fractal dimension, and lyapuna exponent are extracted. The extracted features are represented as $E_{e x t}$ .

3.7. Dimensionality reduction

Afterwards, the extracted features’ dimensions $(E_{e x t})$ are reduced by PCA. The PCA is a significant technique for dimensionality reduction. Through feature organization into a matrix, relationship analysis, eigenvalue and Eigenvector calculations, and data transformation into principal components, Principal Component Analysis (PCA) is a popular technique for reducing dimensionality in EEG signals. Let D the number of feature samples $(ℑ)$ utilized from $E_{e x t}$ , and it is given as,

\begin{aligned} ℑ = {ℑ_{1}, ℑ_{2}, \dots \dots, ℑ_{D}} \end{aligned}

(22)

Next, the mean $(μ)$ and standard deviation $(σ)$ of the samples are represented as,

\begin{aligned} μ & = \frac{\sum ℑ}{D} \end{aligned}

(23)

\begin{aligned} σ & = ℑ - μ \end{aligned}

(24)

After that, the covariance matrix $(ν)$ is generated, and it is depicted as,

\begin{aligned} ν = \frac{1}{D} ℑ^{T r} ℑ \end{aligned}

(25)

Then, the Eigenvalues $(ε)$ and Eigenvectors $(ω)$ are depicted as,

\begin{aligned} ε = ω^{- 1} ν ω \end{aligned}

(26)

Here, $ω^{- 1}$ denotes the inverse of the Eigenvector. Finally, the features’ dimensionality from $E_{e x t}$ is reduced by,

\begin{aligned} ϑ = ω^{- 1} σ \end{aligned}

(27)

Where, $ϑ$ denotes the dimensionality reduced features, and it is depicted as,

\begin{aligned} ϑ = {ϑ_{1}, ϑ_{2}, \dots \dots, ϑ_{d}} \end{aligned}

(28)

Here, d denotes the total number of features. Then these features are given for epilepsy detection.

3.8. Epilepsy classifications

Based on $ϑ$ this, epilepsy is classified by MsRs-LSTM. The LSTM incorporates sophisticated gating mechanisms, which improves the ability of the LSTM to capture complex patterns in sequential data. However, the LSTM is susceptible to training instabilities, such as vanishing gradients, exploding gradients, and overfitting. So, to overcome these issues, the meta-step gradient pruning regularization and rootsig activation function is used. The process of the proposed MsRs-LSTM is described below, and the structure of the proposed MsRs-LSTM is indicated in Fig. 2.

(a). Forget gate: The forget gate $(Γ_{f})$ removes the unused data in the preceding cell state and determines the data passed through the cell, and it is represented as,

\begin{aligned} Γ_{f} = Λ Θ \cdot (ϖ_{f} \cdot [h_{t - 1}, ϑ] + Ω_{f}) \end{aligned}

(29)

Figure 2.

Structure of proposed MsRs-LSTM.

Here, $Λ$ denotes the meta-step gradient pruning regularization, $h_{t - 1}$ denotes the previous hidden state, $Ω_{f}$ represents the bias factor, $ϖ_{f}$ denotes the weight factor, and $Θ$ denotes the rootsig activation, which is depicted as,

\begin{aligned} Θ = \frac{ϑ}{1 + \sqrt{1 + ϑ^{2}}} + 0.5 \end{aligned}

(30)

Then, the meta-step gradient pruning regularization is expressed as,

\begin{aligned} Λ = J^{'} (φ) L^{'} (\hat{φ}) \end{aligned}

(31)

Here, $J^{'} (φ)$ denotes the Jacobian matrix and $L^{'} (\hat{φ})$ represents the gradient vector.

(b). Input gate: The input gate $(Γ_{i})$ controls the new information to flow into the cell state $(Γ_{c})$ , and it is described as,

\begin{aligned} Γ_{i} = Θ \cdot {\tilde{Γ}}_{c} (ϖ_{i} \cdot [h_{t - 1}, ϑ] + Ω_{i}) \end{aligned}

(32)

Here, $Ω_{i}$ and $ϖ_{i}$ denote the bias and weight factors of the input gate, respectively, and ${\tilde{Γ}}_{c}$ denotes the vector function of $Γ_{c}$ .

(c). Cell state: The cell state $(Γ_{c})$ is considered the network's memory, storing the data. Then, ${\tilde{Γ}}_{c}$ and $Γ_{c}$ are represented as,

\begin{aligned} {\tilde{Γ}}_{c} & = \tanh \cdot (ϖ_{c} \cdot [h_{t - 1}, ϑ] + Ω_{c}) \end{aligned}

(33)

\begin{aligned} Γ_{c} & = Γ_{f} \circ Γ_{c - 1} + Γ_{i} \circ {\tilde{Γ}}_{c} \end{aligned}

(34)

Where, $Γ_{c - 1}$ denotes the previous cell state.

(d). Output gate: The output gate $(Γ_{o})$ regulates the information in $Γ_{c}$ and decides the next hidden state $(h_{t})$ , and it is given as,

\begin{aligned} Γ_{o} & = Θ \cdot (ϖ_{o} \cdot [h_{t - 1}, ϑ] + Ω_{o}) \end{aligned}

(35)

\begin{aligned} h_{t} & = Γ_{o} \cdot \tanh (Γ_{c}) \end{aligned}

(36)

Where

h_{t}

denotes the next hidden state. Finally, it

Γ_{o}

is obtained with normal or epilepsy. The pseudocode for DWLSTM is presented as follows,

Thus, the framework effectively detected epilepsy, and the results of the proposed framework are discussed further.

4. Results and discussions

This section analyzes the proposed model's performance by comparing it with the prevailing methodologies for different metrics. The proposed model is executed in the MATLAB platform with wider applicability.

4.1 Dataset description

The Children's Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) Scalp EEG database analyses the proposed system's performance. This dataset comprises the collection of EEG recordings of the 22 pediatric subjects having untraceable seizures. EEG data analyses brain electrical activity to identify seizure clusters and epilepsy. The CHB-MIT Scalp EEG database confirms the suggested framework, acts as a diagnostic tool, and offers information for recognizing seizure patterns. The framework's usefulness is demonstrated by the fact that EEG makes feature extraction for machine learning possible and makes it easier to compare performance with current techniques. Overall, 182 number of seizures are annotated. From the dataset, 80% of the data is utilized for training, whereas the remaining 20% is utilized for testing the proposed model. The dataset link used for the analysis is mentioned below in the reference list. The image results attained by the proposed technique are shown below.

Table 1 shows the image outcome of the normal EEG signal and the epileptic EEG signals.

Table 1.
Image results.

4.2 Performance evaluation

Here, the proposed MsRs-LSTM's performance is analyzed by comparing it with the existing techniques, namely LSTM, Bi-directional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), and CNN (Figure 3).

Figure 3.

Effectiveness evaluation for the proposed MsRs-LSTM.

This figure compares the proposed MsRs-LSTM's effectiveness based on sensitivity, specificity, accuracy, precision, and f-measure with several existing algorithms. Here, the proposed MsRs-LSTM attained 98.78%, 98.75%, 97.56%, 98.67%, and 98.68% of sensitivity, specificity, accuracy, precision, and f-measure, respectively. Thus, due to meta-step gradient pruning regularization integration, the proposed framework enhances the system's performance more than the other prevailing techniques by accurately identifying epileptic seizures while minimizing false positives and negatives.

Table 2 represents the proposed MsRs-LSTM's performance for epilepsy prediction regarding Negative Predictive Value (NPV), False Positive Rate (FPR), False Negative Rate (FNR), True Positive Rate (TPR), and True Negative Rate (TNR). The proposed model attained an NPV of 98.69%, FPR of 0.0068, FNR of 0.0056, TPR of 98.78% and TNR of 98.75%. As the introduced Rootsig activation function augments the learning efficiency of the network, the proposed network detected epilepsy with higher performance. Meanwhile, the existing LSTM attained 95.89% of NPV; Bi-LSTM attained FPR of 0.0304; GRU attained TPR of 88.44%; and CNN attained TNR of 82.44%, which are inferior to the proposed model. Thus, the proposed network's performance is better than the existing techniques.

Table 2.

Performance comparison for MsRs-LSTM.

Methods	NPV (%)	FPR	FNR	TPR (%)	TNR (%)
Proposed MsRs-LSTM	98.69	0.0068	0.0056	98.78	98.75
LSTM	95.89	0.0096	0.0097	94.68	94.67
Bi-LSTM	90.68	0.0304	0.0562	90.58	91.56
GRU	88.69	0.0568	0.087	88.44	89.69
CNN	85.69	0.1564	0.169	86.44	82.44

4.3 Effectiveness analysis for the proposed De-Nf

This section analyses the proposed framework's performance in evaluating the proposed De-NF regarding Structural Similarity Index Measure (SSIM), Mean Square Error (MSE), Mean Absolute Percent Error (MAPE), Root Mean Square Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR) (Figure 4).

Figure 4.

Performance analysis for the proposed De-NF.

Due to the damping error integration, the proposed framework's performance increases regarding SSIM, MAE, MAPE, and RMSE. The performance is evaluated by comparing with traditional methods, namely Notch Filter (NF), Wiener Filter (WF), Gaussian Filter (GF), and Bilateral Filter (BF). This figure indicates that the proposed De-NF achieved 0.9264, 0.0034, 00046, and 0.0045 for SSIM, MAE, MAPE, and RMSE, respectively. Thus, compared to the other conventional techniques, the proposed framework performs better by providing accurate and reliable assessments of image quality and predictive accuracy.

Figure 5 showcases the proposed De-NF's performance based on validating the PSNR. This figure also indicates that the proposed De-NF has high PSNR, which enhances the system performance by minimizing distortion, ensuring superior fidelity, and maintaining high-quality visual information. Moreover, the proposed De-NF attained 65.77 PSNR, whereas the conventional NF attained 61.45 PSNR. Thus, the proposed De-NF outperforms the other prevailing techniques due to the integration of damping error. Using MsRs-LSTM and Fid-DO techniques, the proposed framework for epilepsy detection increases accuracy and reliability over current approaches. It achieves great accuracy and sensitivity by addressing noise problems in EEG signals. The sensitivity of the Damping Error Noise Filter (De-NF) is 98.8%, indicating its effectiveness. Future research should concentrate on the generalisability of datasets and sophisticated methodologies.

Figure 5.

PSNR validation of the proposed De-NF.

Table 3 indicates the comparative evaluation of the proposed work with the prevailing works. The existing works developed models like RNN-LSTM, CNN-RNN, ANN, SVM, and deep forest classifiers for diagnosing focal and generalized epilepsy, epileptic seizure prediction, and ESD. The existing works achieved sensitivity rates like 96.8%, 93.8%, 97.5%, 95.2%, and 98% for epilepsy detection, which are considerably lower than the proposed model. Because the proposed model overcomes the vanishing gradient issue by integrating meta-step gradient pruning regularization and rootsig activation function, the proposed model's sensitivity rate is 98.8%. Hence, the comparative analysis states that the proposed methodology is a significant model for identifying epilepsy.

Table 3.

Comparative analysis with related works.

Author name	Objective	Technique	Sensitivity (%)
Proposed	Epilepsy and seizure clusters detection	MsRs-LSTM	98.8
²³	Diagnosing focal and generalized epilepsy	RNN-LSTM	96.8
²⁴	Prediction of epileptic seizures	CNN and LSTM	93.8
²⁵	Epilepsy Seizure Detection (ESD)	ANN and SVM	97.5
²⁶	ESD	Deep forest	95.2
²⁷	ESD	CNN-RNN	98

Using MsRs-LSTM and Fid-DO techniques, the suggested framework for epilepsy detection has demonstrated notable gains in accuracy and dependability when compared to current approaches. In addition to achieving great sensitivity and accuracy, the model successfully handles EEG signal noise. One significant addition is the thorough preprocessing of EEG data, which includes noise reduction, contrast enhancement, and normalization. High SSIM and low MAE and RMSE readings demonstrate the effectiveness of the Damping Error Noise Filter (De-NF). With a sensitivity of 98.8%, the model outperforms other models, such as RNN-LSTM (96.8%) and CNN-LSTM (93.8%). The model integrates the rootsig activation function with meta-step gradient pruning regularization. This enhancement lessens the vanishing gradient issue frequently arising in deep learning models. One important component for clinical applications in epilepsy management is the ability of the MsRs-LSTM architecture to improve the identification of multiple seizures and seizure clusters.

5. Conclusion

This work developed an effective epilepsy detection framework by detecting the presence of multiple seizures using the proposed Fid-DO and MsRs-LSTM methods. The model preprocessed the EEG signals, by which the noise was removed using De-NF with SSIM of 92.64%, MAE of 0.0034, and RMSE of 0.004. Further, the signal peaks were identified and the features were extracted. Also, the weak periodicity of the signal was analyzed using Fid-DO. Subsequently, the feature dimension was suppressed, and then the epilepsy was detected using the proposed MsRs-LSTM with 97.56% accuracy, 98.67% precision, 98.68% f-measure, and 0.0068 FPR. The implemented results exhibited the enhanced performance of the proposed model over the existing techniques. Hence, a significant detection framework for Epilepsy detection was presented by the proposed methodology.

Future work

Although multiple seizures were detected in this work, the localization of epilepsy was not analyzed. Thus, in future work, the epilepsy localization will be identified with advanced techniques to enhance the epilepsy diagnosis further.

Footnotes

Authors’ contributions

Dr Priyan Malarvizhi Kumar, Dr Wael Korani is responsible for designing the framework, analyzing the performance, validating the results, and writing the article. Tayyaba Shahwar, Dr Gokulnath C is responsible for collecting the information required for the framework, provision of software, critical review, and administering the process.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

No datasets were generated or analyzed during the current study.

References

Spagnoli

Fusco

Percesepe

, et al. Genetic neonatal-onset epilepsies and developmental/epileptic encephalopathies with movement disorders: a systematic review. Int J Mol Sci 2021; 22: 4202.

Xiong

Wang

Zhang

, et al. A study on seizure detection of EEG signals represented in 2D. Sensors 2021; 21: 5145.

Lecce

Smith

Burbach

. Digital mental health interventions for people with epilepsy: a systematic review. Seizure 2023; 107: 91–103.

Rodríguez-Ortiz

Martínez-Torres

. Mutants of the zebrafish K+ channel Hcn2b exhibit epileptic-like behaviours. Int J Mol Sci 2021; 22: 11471.

Milne-Ives

Dunn-Henriksen

Blaabjerg

, et al. At home EEG monitoring technologies for people with epilepsy and intellectual disabilities: a scoping review. Seizure 2023; 110: 11–20.

Ahmed

MIB

Alotaibi

Dash

, et al. A review on machine learning approaches in identification of pediatric epilepsy. SN Computer Science 2022; 3: 437.

Scorrano

Lattanzi

Salpietro

, et al. The cognitive and behavioural effects of perampanel in children with neurodevelopmental disorders: a systematic review. J Clin Med 2024; 13: 372.

Fukuda

Matsuo

Fujimoto

, et al. Vagus nerve stimulation therapy for drug-resistant epilepsy in children—A literature review. J Clin Med 2024; 13: 780.

Mutarelli

Nogueira

Felix

, et al. Modified atkins diet for drug-resistant epilepsy: a systematic review and meta-analysis of randomized controlled trials. Seizure: European Journal of Epilepsy 2023; 112: 77–83.

10.

Adamczyk

Węgrzyn

Wilczyński

, et al. The most common lesions detected by neuroimaging as causes of epilepsy. Medicine (Baltimore) 2021; 57: 294.

11.

Bandopadhyay

Singh

Ghoneim

, et al. Recent developments in diagnosis of epilepsy: scope of MicroRNA and technological advancements. Biology (Basel) 2021; 10: 1097.

12.

Prasanna

Subathra

MSP

Mohammed

, et al. Automated epileptic seizure detection in pediatric subjects of CHB-MIT EEG database—a survey. J Pers Med 2021; 11: 1028.

13.

Ahmedt-Aristizabal

Armin

Hayder

, et al. Deep learning approaches for seizure video analysis: a review. Epilepsy Behav 2024; 154: 109735.

14.

McCallan

Davidson

, et al. Epileptic multi-seizure type classification using electroencephalogram signals from the Temple University Hospital Seizure Corpus: a review. Expert Syst Appl 2023; 234: 121040–121040.

15.

Lemoine

Briard

Rioux

, et al. Computer-assisted analysis of routine EEG to identify hidden biomarkers of epilepsy: a systematic review. Comput Struct Biotechnol J 2023; 21: 4805–4819.

16.

Chiang

Haut

Ferastraoaru

, et al. Individualizing the definition of seizure clusters based on temporal clustering analysis. Epilepsy Res 2020; 163: 1–10.

17.

Wan

, et al. A new localization method for epileptic seizure onset zones based on time-frequency and clustering analysis. Pattern Recognit 2021; 111: 2–39.

18.

Mitsis

Anastasiadou

Christodoulakis

, et al. Functional brain networks of patients with epilepsy exhibit pronounced multiscale periodicities, which correlate with seizure onset. Hum Brain Mapp 2020; 41: 2059–2076.

19.

Deevi

. Developing an integrated machine learning framework for improved brain tumour identification in MRI scans. Current Science & Humanities 2024; 12: 1–18.

20.

Akbarian

Erfanian

. A framework for seizure detection using effective connectivity, graph theory, and multi-level modular network. Biomed Signal Process Control 2020; 59: 1–15.

21.

Supriya

Siuly

Wang

, et al. New feature extraction for automated detection of epileptic seizure using complex network framework. Appl Acoust 2021; 180: 1–12.

22.

Naresh

KRP

. Applying discrete wavelet transform for ECG signal analysis in IOT health monitoring systems. International Journal of Information Technology & Computer Engineering 2022; 10: 157–175.

23.

Zhan

. An epilepsy detection method using multiview clustering algorithm and deep features. Comput Math Methods Med 2020; 2020: 1–11.

24.

Gudivaka

Hussein

Aruna

, et al. Cloud-based Early Acute Lymphoblastic Leukemia Detection Using Deep Learning based Improved YOLO V4. In: 2024 Second International Conference on Data Science and Information System (ICDSIS), 2024, May, (pp. 1–4). IEEE.

25.

Sukriti

Chakraborty

Mitra

. Automated detection of epileptic seizures using multiscale and refined composite multiscale dispersion entropy, chaos. Solitons and Fractals 2021; 146: 1–11.

26.

Najafi

Jaafar

Remli

, et al. A classification model of EEG signals based on RNN-LSTM for diagnosing focal and generalized epilepsy. Sensors 2022; 22: 1–13.

27.

Aslam

Usman

Khalid

, et al. Classification of EEG signals for prediction of epileptic seizures. Applied Sciences (Switzerland) 2022; 12: 1–15.

28.

Omidvar

Zahedi

Bakhshi

. EEG Signal processing for epilepsy seizure detection using 5-level Db4 discrete wavelet transform, GA-based feature selection and ANN/SVM classifiers. J Ambient Intell Humaniz Comput 2021; 12: 10395–10403.

29.

Liu

Wang

Shang

, et al. Epileptic seizure detection based on variational mode decomposition and deep forest using EEG signals. Brain Sci 2022; 12: 1–19.

30.

Malekzadeh

Zare

Yaghoobi

, et al. Epileptic seizures detection in EEG signals using fusion handcrafted and deep learning features. Sensors 2021; 21: 1–28.

A significant framework for epilepsy and seizure clusters detection using MSRS-LSTM and FID-DO

Abstract

Keywords

1. Introduction

1.1. Problem statement

1.2. Objectives

2. Literature survey

3. Proposed methodology for detecting epilepsy using msrs-lstm and fid-do

3.4. Sub-band detection

3.7. Dimensionality reduction

4.1 Dataset description

Table 1. Image results.

Future work

Footnotes

Authors’ contributions

Funding

Declaration of conflicting interests

Data availability statement

References

Table 1.
Image results.