ResneXt-Lenet: A hybrid deep learning for epileptic seizure prediction

Abstract

Encephalopathy is the result of epilepsy, which is defined as recurring seizures. Around the world, almost 65 million people suffer with epilepsy. Because an epileptic seizure involves a crucial clinical element and a clear contradiction with everyday activities, it can be difficult to predict it. The electroencephalogram (EEG) has been the established signal for clinical evaluation of brain activities. So far, several methodologies for the detection of epileptic seizures have been proposed but have not been effective. To bridge this gap, a powerful model for epileptic seizure prediction using ResneXt-LeNet is proposed. Here, a Kalman filter is used to preprocess the EEG signal to reduce noise levels in the signal. Then, feature extraction is performed to extract features, such as statistical and spectral. Feature selection is done using Fuzzy information gain that suggests appropriate choices for future processing, and finally, seizure prediction is done using hybrid ResneXt-LeNet, which is a combination of ResneXt and Lenet. The proposed ResneXt-LeNet achieved excellent performance with a maximum accuracy of 98.14%, a maximum sensitivity of 98.10%, and a specificity of 98.56%.

Keywords

Electroencephalogram (EEG)Kalman filter epileptic seizure LeNet ResneXt

1. Introduction

EEG is an effective symptomatic tool for learning the operational anatomy of the brain. Epilepsy prediction and prescription are studied using EEG, moreover, EEG signals are unstable and non-Gaussian. It is used to measure electrical brain activity and determine the brain disorder types. The analysis of EEG measurements helps to segregate normal and abnormal functions occurring in the brain [1]. The frequency domain and time domain are presented in the EEG signal, these domains arrange two views for a similar data source. Frequency domains illustrate the distribution force of signal concerning the frequency components. Time domains illustrate how the signals change over time [2]. During a normal situation, EEG signal patterns are common to normal subjects. Moreover, the patterns of action potentials are suddenly transformed and the EEG signal comes in as fast spike waves. The patterns of the EEG signals are categorized into four methods; ictal method, preictal method, postical method and interictal method [3]. The EEG signal prototypes are specific to patients with exclusive epilepsy. Hence, the epilepsy patient-specific seizure prediction gives better accuracy compared to patient-independent seizure prediction [4]. If the patient has drug resistance, it is usually treated with either anti-epileptic drugs or possibly surgical procedures to remove the affected part of the brain [5]. Experts are even focused on finding successful therapies for seizure management. Therefore, early prediction of seizures has become very important, as it ensures sufficient time to use an effective medication to treat or eliminate the seizure [6].

Each year epilepsy affects 2.4 million people of all ages worldwide and presents seizures with the risk of periodic disruptions in cognitive and behavioral functions [7]. Epilepsy is the result of the simultaneous sudden filtering of brain neurons. EEG is used to detect the electrical activity of the brain. Seizures are a minor multiple number of channels for EEG recording. Generalized seizures are viewed in every channel of the EEG recording [8, 9]. The patient with epilepsy disease is affected critically whenever a seizure occurs as a result of uncontrollable movements, loss of consciousness and convulsions that can lead to serious injuries and sometimes the people’s death [10, 11, 12]. A Discrete wavelet transform (DWT) is a method in which ripple transforms are captured and the signal is individually sampled to find different signal properties of the signal. Frequency and location information are captured through DWT. The different techniques are wavelet series, wavelet transform, and wavelet compression [13, 14]. The seizure prediction procedure can be formulated as a classification problem between EEG signals to distinguish the interictal brain state from the preictal state [6]. As a result, the problem of predicting seizures can be defined as a duplex categorization problem where label value zero indicates the interictal phase and label one represents the preictal stage. Seizure prediction is more challenging and useful than seizure detection since seizure prediction allows the patients or medical staff to take appropriate actions to avoid the occurrence of seizures [2].

The recent developments of deep learning (DL) techniques have been established to perform traditional machine learning systems in a wide variety of disciplines, such as computer vision, speech recognition, natural language processing (NLP), and biomedical applications [15]. Various DL structures are used for the categorization of EEG brain signals. Shallow artificial neural networks (ANNs) used to be applied only as classifiers after traditional feature extraction while convolutional neural networks (CNN) have been utilized for both automatic feature learning and classification [16, 17, 18, 19, 20, 6]. Due to the availability of large data sets, approaches are fixed in DL techniques are prepared which significantly improves the determination of different neurological disorders, as well as epileptic seizures. The DL-based CAD systems enable physicians to make better-informed decisions based on the recorded patient neuroimaging modalities [21]. Recently, DL methods, such as CNN with automatic retrieval of relevant information are very popular functions. DL models are used in various aspects, such as speech recognition and computer vision. DL models can extract spectral features, and the EEG epilepsy data contains temporal features that can be utilized to understand the overall pattern of a seizure. This method can serve as a powerful multimedia tool for generating fast results and brain mapping images available for medical experts to use in their further investigations [22].

The main aim of this paper is to design ResneXt-LeNet to achieve epileptic seizure prediction using EEG signals. The input EEG signal is first preprocessed and put into feature extraction to extract features, such as statistical and spectral. Then a fuzzy information retrieval model is presented to select suitable features for superior prediction. Finally, epilepsy seizure is predicted using ResneXt-LeNet, which is the combination of ResneXt and Lenet.

The major contribution of this paper is represented by,

•
Proposed ResneXt-Lenet for epilepsy seizure prediction: The prediction of an epileptic seizure is achieved by the ResneXt-Lenet model, which is a combination of the ResneXt and Lenet models.

The rest of this paper is sorted by: The analyzed works along with their strengths and limitations are presented in Section 2, the proposed ResenXt-LeNet for epileptic seizure prediction is explained by schematic structures in Section 3, results for ResenXt-LeNet are presented in Section 4 and final part is given in Section 5.
2. Motivation

Epilepsy is a chronic nervous ailment of the brain that affects all age groups of people. It is categorized by erratic seizures that affect the patient’s mental health and frequently lead to depression, anxiety, or cognitive impairment. EEG is a dimension of the electrical impulses caused by brain cells. If the patient does not take appropriate treatment properly, there is a decrease in the chances of survival, which leads to sudden death. Hence, the early prediction of epilepsy seizure is important and it is a challenging process. Therefore, an early stage of epileptic seizure prediction is necessary for an innovative model.

2.1 Literature survey

Ibrahim SW et al. [23] introduced the K-nearest neighbor (KNN) method for continuously monitoring the EEG signals by comparing the current sliding window with normal and pre-seizure baselines to predict seizure. It can be efficiently implemented in any implanted system that is mobile devices or Field-Programmable Gate Arrays (FPGA) board. However, it did not take a larger dataset to test this method. Hussein R et al. [24] designed the Deep Neural Network (DNN) to determine the temporal dependencies in EEG data for strong detection of epileptic seizures. It achieved seizure detection attention to deprivation conditions. However, it only addressed single-channel EEG data. Liu CL et al. [2] introduced the Multi-view CNN to predict the incidence of epilepsy seizures to acquire a shared representation of time-domain and frequency-domain features. This method outperformed the methods in the leaderboard of the Kaggle competition. However, this method was not applicable in the brain-computer interface and many other areas. Inoue M et al. [25] modeled the Deep Recurrent Neural Network (DRNN) for human activity recognition with high throughput from raw accelerometer data. It secured a high acceptance rate and throughput. However, the recognition rate of the sequence data was decreased. Jana R and Mukherjee I [4] introduced CNN for automated extraction of features and classification of epilepsy patient’s states. The raw EEG signals in this method are utilized for automated feature extraction. Patient-specific seizure prediction is unable to effectively predict seizures in other epilepsy patients. Dissanayake T et al. [26] designed a DL classifier for patient-independent epileptic seizure prediction. It secured high inter-subject variability. Although, designing a classifier is a complex task to generalize completely invisible subjects. Jemal I et al. [7] introduced an Interpretable DL model for seizure prediction using EEG signals. It achieved a fairly high level of prediction accuracy, but it is a complex task for the development of patient-independent models. Abdelhameed AM and Bayoumi M [6] designed a Novel Deep Learning system for epileptic seizure prediction using multi-channel EEG recordings from the scalp of human brains. Within one hour, this method successfully forecast 52 out of 55 seizures. However, it failed to achieve a high average false prediction rate.

2.2 Challenges

Major problems undertaken during epilepsy seizure prediction using hybrid deep learning are listed as follows,

•
In [4], CNN was devised for automatic symptom extraction and status classification of patients with epilepsy. However, this method failed to implement a patient-independent seizure prediction model as it works efficiently for all epilepsy patients.
•
Although the multi-view CNN in [2] predicted occurrences of seizures, it failed to design a recent architecture for the Electrocardiography (ECG) test, it is the process of recording the electrical activity of the heart using electrodes for a specified period and it placed on the skin.
•
KNN was employed to continuously monitor the EEG signals but it was unsuccessful to enhance prediction accuracy with a robust feature extraction method [23].
•
In [24], DNN was devised for automated recognition of epileptic seizures using EEG signals and it is effective in high-level EEG images and could properly be distinguished between the seizure EEG activities. However, it did not fully manipulate the ability of deep neural networks in seizure detection in terms of the detection power of brain data time series.
•
Epilepsy is a neurological disease characterized by unconsciousness and convulsion. Epileptic patients suffer from epileptic seizures due to anomalous electrical discharges. It is essential to improve the quality and accuracy of captured data. On the other hand, inaccurate documentation of seizures negatively affects patient care.

3. Proposed ResneXt-Lenet for epileptic seizure prediction

Epilepsy is a nervous disorder that affects strange behavior, feelings, and loss of awareness, so the earlier diagnosis of epileptic seizure is very important for effective treatment and decreasing the risk of injury. Consequently, this research proposes an effective model for epileptic seizure prediction using a newly developed methodology named ResneXt-Lenet. The overall process includes preprocessing, feature extraction, feature selection and epileptic prediction. The EEG signal is preprocessed by the Kalman filter [27] to reduce the noisy content present in the EEG signal. Subsequently, feature extraction is executed, where features like statistical and spectral features are extracted. Then, the selection process is done using fuzzy information gain [28] to select the appropriate feature. At last, the epileptic prediction is done using the ResneXt-LNet model. Figure 1 displays the block diagram of ResneXt-Lenet for epileptic seizure prediction.

Figure 1.

Function block diagram of ResneXt-Lenet for epileptic seizure prediction.

3.1 Image acquisition

Consider the dataset $G$ having $m$ number of signals that are mathematically described as follows,

$\displaystyle G=\{{R_{1},R_{2}\ldots R_{m}\ldots R_{n}}\}$ (1)

Here, $R_{m}$ represents the $m^{\text{th}}$ number of signals present in the database $G$ and $R_{n}$ denotes the total number of signals

3.2 EEG signal preprocessing using Kalman filter

The initial step in predicting an epileptic seizure is pre-processing. Here, the entering signal $R_{m}$ is given into the pre-processing module for the removal of noisy content in the carrier signal using the Kalman filter. These signals are obtained from the brain. Due to spatial information loss, it failed to provide an accurate signal. Therefore, the pre-processing step is very suitable for eliminating noisy content. So, high-quality signals are generated for increased prediction performance. Kalman filters are the most widely used variant of Bayes filters. The mathematical representation is described below,

$\displaystyle Bel({x^{\prime}_{t_{1}}})=\frac{1}{(2)^{\frac{d_{1}}{2}}|{\sum% \nolimits_{t}}|^{\frac{1}{2}}}\exp\left[{\frac{-1}{2}({x^{\prime}_{t_{1}}-\xi_% {t_{1}}})^{T_{1}}\sum\nolimits_{t_{1}}^{-1}{({x^{\prime}_{t_{1}}-\delta_{t_{1}% }})}}\right]$ (2)

where $\delta_{t_{1}}$ denotes distribution mean, $\sum\nolimits_{t_{1}}$ represents uncertainty if the covariance is the larger and wider the distribution’s spread. Moreover, the pre-processed EEG signal is labelled $P_{m}$ , and t $M\times N$ is the dimension size.

3.3 Feature extraction

The pre-processed $P_{m}$ signal is subject to a feature extraction stage where the feature extraction is performed based on spectral and statistical features.

3.3.1 Statistical features

It effectively separates EEG signals from both scalp EEG and intracranial signals. In addition, statistical features extracted from the pre-processed $P_{m}$ signal are given below.

i) Mean

Mean values are estimated by taking the average of total samples that exist in the EEG signal, which is denoted by

$\displaystyle A_{1}=\frac{1}{\beta}\sum\limits_{s=1}^{\beta}{P_{m}}$ (3)

where, $P_{m}$ is the preprocessed signal $\beta$ indicates the entire samples and $A_{1}$ represents the mean feature.

ii) Variance

Variance is established on the mean, which is estimated using the $n$ number of samples and is defined by,

$\displaystyle A_{2}=\frac{1}{\beta}\sum\limits_{s=1}^{\beta}{({P_{m}-A_{1}})^{% 2}}$ (4)

where, $A_{1}$ denotes the mean feature and $A_{2}$ is represented by the variance feature.

iii) Skewness

Skewness is utilized to provide the information concerning data symmetry, and the skewness is computed using the below expression,

$\displaystyle A_{3}=\frac{1}{\beta}\sum\limits_{s=1}^{\beta}{({P_{m}-A_{1}})^{% 3}}$ (5)

where, $A_{3}$ denotes the skewness feature.

iv) Kurtosis

Kurtosis is used to obtain information about peaks in the signal. The output of the kurtosis feature is given below,

$\displaystyle A_{4}=\frac{1}{\beta}\sum\limits_{s=1}^{\beta}{({P_{m}-A_{1}})^{% 4}}$ (6)

where, $A_{4}$ denotes the kurtosis feature.

Thus, the pre-processed signal is represented as $F_{1}$ and it is given by,

$\displaystyle F_{1}=\{{A_{1},A_{2},A_{3},A_{4}}\}$ (7)

3.3.2 Spectral features

This feature is also known as frequency domain features, where, features are extracted from EEG signals under scalp EEG signals. The spectral features are described below,

i) PSD

PSD measures the power of EEG signal frequency function and the mathematical expression is given below,

$\displaystyle f_{1}=\sum\limits_{s=1}^{\beta}{\alpha({P_{m}})}_{m^{-\textit{% bjp}}}$ (8)

Where, $\alpha({P_{s}})$ represents autocorrelation of preprocessed signal, $f_{1}$ represents PSD feature.

ii) Spectral centroid

The information about variation in signal is estimated using spectral centroid, and the equation is defined by,

$\displaystyle f_{2}=\frac{\sum\nolimits_{p}{R({f_{1}})}}{f_{1}}$ (9)

where, spectral centroid features are denoted by $f_{2}$ and $p$ indicates the EEG sub-band.

iii) Variational coefficients

Variational coefficients are said to be a frequency band property that calculates the deviation in EEG recording, and the equation expressed below,

$\displaystyle f_{3}=\sqrt{\frac{\sum\nolimits_{p}{({R-f_{2}})^{2}.({f_{1}})}}{% f_{1}}}$ (10)

where, $f_{3}$ represents the variational coefficient feature.

iv) Spectral skewness

Spectral skewness refers to the spectral characteristics that show strong performance in scalp EEG signals and it is expressed as,

$\displaystyle f_{4}=\frac{\sum{({({R-f_{2}})/f_{3}})^{3}.({f_{1}})}}{f_{1}}$ (11)

Hence, the extracted spectral features from the pre-processed signal are denoted by $F_{2}$

$\displaystyle F_{2}=\{{f_{1},f_{2},f_{3},f_{4}}\}$ (12)

3.3.3 Logarithmic band power

It is derived from the logarithmic power and different EEG signal bands. This feature was determined by calculating the power value of the frequency band and it is expressed below,

$\displaystyle g_{1}=\log({P_{m}})$ (13)

where, logarithmic band power is termed as $g_{1}$ .

3.3.4 CSP

The key concept is the use of linear conversion that converts multichannel EEG assessment into a lower-level dimensional spatial subgroup using the projection matrix. The weights of each channel comprise each row of the projection matrix. Thus, a common spatial pattern is expressed as,

$\displaystyle g_{2}=\log\left[{\frac{A_{2}({\lambda_{y}})}{\sum\limits_{s=1}^{% y}{A_{2}({\lambda_{s}})}}}\right]$ (14)

where, $\lambda y$ denotes the $y^{\text{th}}$ row of $\lambda$ vector, and $g_{2}$ indicates the CSP feature.

3.3.5 Wavelet coefficient

Wavelets refer to sharp waves containing zero mean values. This feature comprises both frequency and time domain of localization capability and the equation becomes,

$\displaystyle g_{3}=\frac{1}{\sqrt{t_{1}}}\mu\left({\frac{\kappa-\gamma}{t_{1}% }}\right)$ (15)

where, $t_{1}$ represents the scaling factor, $\gamma$ denotes the shifting factor, and the term $g_{3}$ represents the wavelet coefficient feature.

3.3.6 Pitch chroma

It is very powerful for effectively classifying the pitches and bringing them closer to an equal moderated pitch scale. For feature extraction, the Fourier transform frequency is obtained that is mapped with 12 semitone classes of pitch (chroma) $C$ . However, the pitch chroma is utilized to minimize the transient noise available in the signals. Consequently, the $P$ vector is determined by mapping chroma and their related semi-tone pitches, and is expressed by,

$\displaystyle C(P)=\sum\limits_{g^{\prime},b({g^{\prime}=p})}{D({g^{\prime}})}% ;P\in[{0,12}]$ (16)

where, the half frequency range is estimated by expanding the Fourier transform of pre-processed signal and the equation is given below.

$\displaystyle D({g^{\prime}})=\sum\limits_{m}{u_{g^{\prime}}}(m)Z(m)$ (17)

where, the Fourier transform value is denoted as $Z(m)$ , the prevalence of Fourier transform was indicated as $m$ . The filter with parameter $y$ is defined as the relative distance between the center of frequencies and the filter of the spectrum.

$\displaystyle u_{g^{\prime}}(m)=\frac{1}{2}\tan j({\sigma({1-2y})})+\frac{1}{2}$ (18)

where, $y=p|{g^{\prime}-g(m)}|$ . Therefore, the output obtained from the pitch chroma feature is $g_{4}$ .

3.3.7 Spectral decrease

The spectral decrease is employed to quantify the reduction of the magnitude spectrum as time progresses. It indicates how much the spectrum decreases when the slopes of lower frequencies increase and it is given as,

$\displaystyle g_{5}=\frac{\sum\limits_{b=P_{1}+1}^{P_{2}}{\frac{D_{b}-D_{b_{1}% }}{b-1}}}{\sum\limits_{B=p_{1+1}}^{P_{2}}{D_{b}}}$ (19)

3.3.8 Spectral flux

EEG signal of spectral information is identified using the spectral flux feature. This information is obtained by utilizing frequency-enabled parameters, and the difference in value between consecutive spectral frames is taken into account. The mathematical expression of spectral flux is denoted as,

$\displaystyle g_{6}=\sum\limits_{R=0}^{\frac{P}{2}-1}{({|{B_{s}(R)}|^{2}-|{B_{% s-1}(R)}|^{2}})}$ (20)

where, the term $B_{s}(R)$ is the EEG signal spectrum value.

3.3.9 Tonal power ratio

It calculates the inflection of preprocessed EEG signals. It refers to the ratio of the hormonal intensity as well as the signal segment. Here, $X(q_{1},r_{1})$ signifies the spectrum of input EEG signal and thus, the tonal power ratio is expressed as,

$\displaystyle g_{7}=\frac{Y({r_{1}})}{\sum\limits_{\omega=0}^{\frac{j}{2}-1}{|% {X({q_{1},r_{1}})}|^{2}}}$ (21)

where, the term $Y(r_{1})$ refers to the tonal power. The output of the tonal power ratio feature is denoted as $g_{7}$ and the overall extracted features are given by,

$\displaystyle F_{3}=\{{g_{1},g_{2},\ldots g_{7}}\}$ (22)

The feature vector can be expressed by $F_{m}$ and the expression can be written as,

$\displaystyle F_{m}=\{{F_{1},F_{2},F_{3},F_{4}}\}$ (23)

The magnitude of the dimension is $[{m\times n}]$ . The extracted feature $F_{m}$ is taken as input for selecting the features for seizure prediction performance.

3.4 Feature selection using fuzzy information gain model

Feature selection is performed by employing a fuzzy information gain method for processing an epileptic seizure prediction. Here, a fuzzy informative gain model is presented in Class Information Gain (CIG), Information Gain (IG), and Mutual Information. Each other information is represented as $\mu$ . Assume another feature $c_{1}$ , so its probability values are denoted as $V({g_{1}})$ and $V({c_{1}})$ . Thus, the mutual information equation is given below,

$\displaystyle\mu=\log_{2}\frac{v({g_{1},c_{1}})}{v({g_{1}})v({c_{1}})}$ (24)

where, $\mu$ denotes the mutual information, $v({g_{1}})$ and $v({c_{1}})$ indicates the feature probability of $gg_{1}$ and $v_{1}$ and $v({g_{1},c_{1}})$ represents joint probability of features $g_{1}$ and $v_{1}$ . The IG feature is expressed as,

$\displaystyle E=\textit{prob}({E_{a}})\log_{2}({\textit{prob}({E_{a}})})$ (25)

where, $E$ is the IG feature contributed by,

$\displaystyle Z=\textit{prob}({V_{x}=1,E_{a}})\log\frac{\textit{prob}({V_{x}=1% ,E_{a}})}{\textit{prob}({V_{x}=1})\textit{prob}({E_{a}})}{}\!+\textit{prob}({V% _{x}=0,E_{a}})\log\frac{\textit{prob}({V_{x}=0,E_{a}})}{\textit{prob}({V_{x}=0% })\textit{prob}({E_{a}})}$ (26)

Where, $x$ denotes the feature, $E_{a}$ represents the category of feature, $V_{x}$ indicates the value of a feature $x$ , $\textit{prob}({v_{x},E_{a}})$ is the possibility of feature $x$ deception in class $E_{a}$ and $Z$ denotes CIG. A fuzzy system takes $\mu,H,Z$ as the inputs and processes the information gain features to fuzzy score, the mathematical expression is given below,

$\displaystyle P_{1}^{\textit{score}}=\omega_{1}({\mu,H,Z})$ (27)

where, $\omega_{1}$ indicates the fuzzy system, $P_{1}^{\textit{score}}$ denotes fuzzy score value. The output of the feature selection is referred to $N_{m}$ and the dimension is $[{m\times l}]$ .

3.5 Epileptic seizure prediction using ResneXt and Lenet

The proposed network introduced for epileptic seizure prediction in this research is named ResneXt-Lenet, which is a combination of ResneXt and Lenet. The common process in this framework is explained as follows. At first, the input signal is passed to the ResneXt model and it generates an output $O_{1}$ . By considering the selected feature $F_{m}$ with $O_{1}$ as input is subjected to ResneXt-Lenet layer for the subsequent process. In this proposed layer, two significant functions are carried out, such as fusion and regression. The output of ResneXT-LeNet is denoted as $O_{2}$ and finally, it is fed into the Lenet model. The outcome received through this proposed hybrid classifier is specified as $O_{3}$ . Figure 2 illustrates the general structure of the ResneXt-LeNet model.

Figure 2.

The general outline of ResneXt-Lenet model.

3.5.1 ResneXt model

The ResneXt block [29] follows the split-transform-merge strategy and it performs a set of transformations as opposed to the ResNet block. Convolution layer outputs from all groups are secured via $1\times 1$ . The magnitude of the transform set is cardinality. Since the ResNeXt block shares the aggregate topology, replaces the rest of the conventional ResNet block where as it is easy to assemble and produce the output $O_{1}$ . The mathematical expression of the ResneXt [30] model is expressed as,

$\displaystyle O_{1}=R_{m}+\sum\limits_{i=1}^{Q}{U_{i}({R_{m}})}$ (28)

Where, $T$ denotes individual transformation, the carrier signal is $R_{m}$ . Figure 3 demonstrates the structure of the ResneXt model.

Figure 3.

The architecture of the ResneXt model.

3.5.2 ResneXt-Lenet Layer

In ResNet, the two dimensions for tuning model capacity are depth and width. Typically, a model with more arguments has a stronger representational ability. ResNeXt is a building block that aggregates a set of transformations with a similar topology. Compared to ResneXt it reveals a new dimension cardinality, which is the set of conversions. In Lenet model consists of a convolutional layer, pooling max layer, ReLu and a Fully activation layer connected layer. Each layer in this network uses the ReLu activation function. An FC layer is present after the pooling layer, which provides a classified result. The output $O_{1}$ and $F_{m}$ is subjected to the ResneXt-Lenet layer and it produces an output $O_{2}$ . Hence, the mathematical expression is given below,

$\displaystyle y_{1}=\sum\limits_{j=1}^{4}{F_{2}\times w_{j}}$ (29)

where, $F_{2}$ refers to the spectral features, $w$ denotes the weight. Then applying the statistical features, the equation becomes

$\displaystyle y_{2}=\sum\limits_{j=1}^{11}{\sum\limits_{k}^{n_{1}}{F_{1}w_{jk}}}$ (30)

where, $F_{1}$ denotes the statistic features, $w$ refers to weight, $y_{1}$ denotes output from statistical features, $y_{2}$ denotes output from statistical features.

By applying the fractional calculus concept, the equation becomes,

$\displaystyle x_{1}({t+1})=h_{1}x_{1}(t)+\frac{1}{2}h_{1}x_{1}({t-1})+\frac{1}% {6}({1-h_{1}})x_{1}({t-2})$ (31)

The mathematical equation of output $O_{2}$ becomes,

$\displaystyle O_{2}=h_{1}y_{1}+\frac{1}{2}h_{1}y_{2}+\frac{1}{6}({1-h_{1}})% \times O_{1}$ (32) $\displaystyle O_{2}=\sum\limits_{j=1}^{4}{F_{1j}\times w_{j}}+\frac{1}{2}h_{1}% \sum\limits_{j=1}^{11}{\sum\limits_{k=1}^{e}{w_{jk}+\frac{1}{6}}}({1-h_{1}})O_% {1}$ (33)

3.5.3 Lenet model

LeNet [31] was developed by LeCun in 2015 and was initially used in various banks to identify handwritten digits on a check. LeNet essentially contains a pooling max layer, convolutional layer, ReLu activation layer and fully connected layer. In LeNet, input data is distributed to a convolutional layer with a kernel filter, while low-level features are excavated. In the LeNet model, the output layer is known as the softmax layer. Moreover, the selected output secures better probability. The mathematical equation of the softmax function is given below,

$\displaystyle O_{3}=\textit{soft}\max({\vartheta^{T}c_{e}^{T-1}+b^{T}})$ (34)

where, $O_{3}$ stands for the final result of the LeNet network. The feature selection process is the purpose of the pooling layer. Whereas it reduces the dimensionality of the data while preserving basic analytical properties. Mean pooling, pooling max and random pooling are general-use advances that extract with median value, largest value, and random values local host domain. The activation layer is assumed to perform a feature recruitment process and reduce data dimensions while maintaining the salient features of the data.

$\displaystyle c_{e}^{i}=\textit{pool}({c_{e}^{i-1}})$ (35)

where, pooling max operation is symbolized as $\textit{pool}(\bullet)$ . Commonly, the output of $i^{\text{th}}$ layer is indicated as $c_{e}^{i}$ and former layer output is $c_{e}^{i-1}$ , $e$ is answering to the $e^{\text{th}}$ sample. Figure 4 shows the architecture of the LeNet model.

Figure 4.

The architecture of the LeNet model.

4. Results and discussion

This segment delineates an outcome of devised ResneXt-LeNet by various evaluation metrics and assesses the performance to prove the superiority level.

4.1 Experimental setup

ResneXt-LeNet model for epilepsy seizure prediction is successfully executed in MATLAB tool.

4.2 Experimental results

The sample image output of the ResneXt-LeNet method for epilepsy seizure prediction is given in Fig. 5. The input EGG signals 1, 2, and 3 are denoted in Fig. 5a, c, and e. The filtered EEG signals 1, 2, and 3 are denoted in Fig. 5b, 5 d, and 5 f.

Figure 5.

Experimental outcomes, a. Input EEG signal-1, b. filtered output of EEG signal-1, c. Input EEG signal-2, d. filtered output of EEG signal-2, e. Input EEG signal-3, f. filtered output of EEG signal-3.

4.3 Dataset description

The dataset used by the ResneXt-LeNet method for epilepsy seizure prediction is CHB-MIT Scalp EEG [32]. It is a group of EEG recordings of 22 child subordinates with unsolvable seizures. The recordings are grouped into 23 cases gathered from 22 subordinates based on data of 5 men and 17 women aged from 3–22 years and 1.5–19 years.

4.4 Evaluation metrics

The effectiveness of the ResneXt-LeNet method for brain epilepsy seizure prediction is evaluated with the three evaluation metrics given below.

4.4.1 Accuracy

It calculates the rate of detection results that are correctly classified and formulated as,

$\displaystyle\textit{Accuracy}=\frac{B^{p}+B^{n}}{B^{p}+B^{n}+R^{p}+R^{n}}$ (36)

where, $R^{p}$ represents a false positive, $B^{p}$ represents a true positive, $R^{n}$ represents a false negative and $B^{n}$ represents a true negative.

4.4.2 Sensitivity

It is the proportion of positive standards that are accurately determined by the classifier and it is described as,

$\displaystyle\textit{Sensitivity}=\frac{B^{p}}{B^{p}+R^{n}}$ (37)

4.4.3 Specificity

Specificity is the proportion of negative impacts that are accurately identified by the classifier and it developed as,

$\displaystyle\textit{Specificity}=\frac{B^{n}}{B^{n}+R^{p}}$ (38)

4.5 Performance analysis

The competitive analysis of the proposed ResneKt-LeNet using different training data. Performance metrics like accuracy, sensitivity and specificity are explained in Fig. 6. The performance analysis of the proposed ResneKt-LeNet using accuracy is shown in Fig. 6a. The accuracy values computed by ResneXt-LeNet with iteration 20, 40, 60, 80, and 100 are 0.770, 0.797, 0.833, 0.869, and 0.888 while the training percentage is 50. When the training data is 90%, the accuracy values of proposed ResneXt-LeNet with iterations 20, 40, 60, 80, and 100 are 0.788, 0.825, 0.862, 0.907, and 0.953. Figure 6b illustrates the performance analysis of sensitivity. When the training data is 90%, the values of sensitivity calculated with iterations 20, 40, 60, 80, and 100 are 0.798, 0.838, 0.875 0.917, and 0.963. The sensitivity of the proposed ResneKt-LeNet is 0.770, 0.811, 0.846, 0.880, and 0.899 for the training data of 50% with the iteration from 20, 40, 60, 80, and 100, respectively. Performance analysis of specificity is shown in Fig. 6c. If the training data is 90%, the specificity values computed with iteration 20, 40, 60, 80, and 100 are 0.794, 0.838, 0.870, 0.917, and 0.974. The specificity of the proposed ResneKt-LeNet is 0.774, 0.800, 0.840, 0.873, and 0.896 when the training data is 50% with the iteration varied from 20, 40, 60, 80, and 100.

Figure 6.

Performance analysis of ResneKt-LeNet, a. Accuracy, b. Sensitivity, c. Specificity.

4.6 Comparative methods

The competitive methods used for analyzing the performance of the ResneKt-LeNet method for epilepsy seizure prediction are KNN [23], NN [24], DeepCNN [2], DeepRNN [25], SASO-based DeepRNN, and EXP-SASO-based DeepRNN.

4.6.1 Analysis based on training data

Figure 7 describes the ResneKt-LeNet analysis with different evaluation scales by the change in training percentage. The analysis of ResneKt-LeNet using accuracy is demonstrated in Fig. 7a. If training data is 40%, the accuracy of KNN is 79.485, NN is 80.490, DeepCNN is 89.924, DeepRNN is 90.344, SASO-based DeepRNN is 91.264, EXP-SASO-based DeepRNN is 92.198, and ResneKt-LeNet is 94.436. Hence, the improved performance of the devised scheme with existing implementations is 15.832%, 14.767%, 4.778%, 4.333%, 3.359% and 2.369%. Figure 7b illustrates ResneXt-LeNet analysis using sensitivity. While the training percentage is 90, the sensitivity of accessible frameworks and ResneKt-LeNet are 83.319, 87.344, 92.370, 93.870, 96.520, 97.506, and 98.100. Moreover, the improved performance while comparing the devised model with comparative techniques is 15.067%, 10.965%, 5.841%, 4.312%, 1.611%, and 0.606%. The designed approach using specificity is illustrated in Fig. 7c. When training data is 90%, the specificity of convolutional methods is 80.059, 82.656, 95.216, 95.376, 97.536, 98.051 and ResneKt-LeNet is 98.555. The performance enhancement value of the designed scheme with KNN is 18.767%, NN is 16.132%, DeepCNN is 3.388%, DeepRNN is 3.225%, SASO-based DeepRNN is 1.034%, EXP-SASO- based DeepRNN is 0.511%.

Figure 7.

Comparative assessment by changing using training data a. Accuracy, b. Sensitivity, c. Specificity.

Table 1

Comparative discussion

Dataset	Metrics/ methods	KNN	NN	DeepCNN	DeepRNN	SASO-based DeepRNN	EXP-SASO-based DeepRNN	ResneXt- Lenet
Training data	Accuracy	82.07%	83.18%	92.10%	93.22%	96.55%	97.52%	98.14%
$=$ 90%	Sensitivity	83.32%	87.34%	92.37%	93.87%	96.52%	97.51%	98.10%
	Specificity	80.06%	82.66%	95.22%	95.38%	97.54%	98.05%	98.56%
K-Fold $=$ 9	Accuracy	79.42%	89.65%	91.03%	92.45%	94.13%	95.11%	98.27%
	Sensitivity	80.24%	91.32%	92.10%	94.47%	96.52%	97.51%	98.05%
	Specificity	82.41%	85.27%	89.02%	92.47%	93.74%	94.73%	98.06%

Figure 8.

Comparative evaluation of ResneKt-LeNet, a. Accuracy, b. Sensitivity, c. Specificity.

4.6.2 Analysis based on K-Fold

Figure 8 describes the ResneKt-LeNet analysis with different K-fold values concerning certain assessment measures. Accuracy is considered in Fig. 8a. When K-Fold is 4 then, the accuracy of convolutional methods is 76.513, 86.982, 87.309, 90.450, 90.798, 91.753, and 93.457. The improved performance achieved when compared with the previous models, such as KNN is 19.184%, NN is 8.773%, DeepCNN is 7.368%, DeepRNN is 5.924%, SASO-based DeepRNN is 4.213%, and EXP-SASO-based DeepRNN is 3.209%. Figure 8b illustrates sensitivity. For K-fold is 5, the sensitivity of KNN is 77.939, NN is 88.635, DeepCNN is 89.870, DeepRNN is 91.699, SASO-based DeepRNN is 92.438, EXP-SASO-based DeepRNN is 93.412, and ResneKt-LeNet is 94.457. The performance gain of the designed scheme in comparison with KNN is 17.487%, NN is 6.164%, DeepCNN is 4.856%, DeepRNN is 2.920%, SASO-based DeepRNN is 2.138%, and EXP-SASO-based DeepRNN is 1.106%. The specificity analysis is shown in Fig. 8c. The existing methods and ResneKt-LeNet gained the specificity of 82.414, 85.274, 89.016, 92.470, 93.736, 94.725, and 98.057, the performance gain values are 15.953%, 13.036%, 9.220%, 5.697%, 4.407% and 3.398 % while K-Fold is 9.

4.7 Comparative discussion

A comparative discussion consisting of the ResneXt-LeNet is shown in Table 1. This analysis is performed using K-Fold and training data established on the metrics functions. From this analytical part, it is clear that experimenting with the use of training data achieved a maximum accuracy of 0.981%, sensitivity of 0.981%, and specificity of 0.985%. Here, the ResneXt-LeNet model achieved excellent performance in taking into account training data as 90%.

5. Conclusion

This work presented ResneXt-LeNet for predicting the epileptic seizures located on EEG signals. Epilepsy is the most common neurological ailment and it occurs in people of all ages. It occurs unexpectedly and thus increases the mortality rate of people. The various steps that ResneXt-LeNet follows are pre-processed using a Kalman filter. Feature extraction that extracts statistical and spectral features, and effectively extracts Spectral decrease, Pitch Chroma, Tonal Power Ratio and Spectral Flux. Feature selection is performed using a fuzzy information gain model to select significant properties. The selected best features are subjected to the seizure prediction model and it is accomplished using ResNeXt-LeNet, which is modelled by a combination of ResneXt and LeNet. The experimentation of the developed ResneXt-LeNet is performed on the CHB-MIT Scalp EEG dataset and effective performance is achieved with a maximum accuracy of 98.14%, sensitivity of 98.10%, and specificity of 98.56%. In future, better prediction performance can be attained by extracting some additional extraction features, and also incorporating Transfer Learning (TL) would help in solving the problem of low detection accuracy.

References

Rasheed

Qayyum

Qadir

Sivathamboo

Kwan

Kuhlmann

, et al. Machine learning for predicting epileptic seizures using EEG signals: A review. IEEE Rev Biomed Eng. 2020; 14: 139-155.

Liu

Xiao

Hsaio

Tseng

. Epileptic seizure prediction with multi-view convolutional neural networks. IEEE access. 2019; 7: 170352-170361.

Elgohary

Eldawlatly

Khalil

. Epileptic seizure prediction using zero-crossings analysis of EEG wavelet detail coefficients. IEEE (CIBCB). 2016; 1-6.

Jana

Mukherjee

. Deep learning based efficient epileptic seizure prediction with EEG channel optimization. Biomed Signal Proces. 2021; 68: 102767.

Elger

Schmidt

. Modern management of epilepsy: A practical approach. Epilepsy Behav. 2008; 12(4): 501-539.

Abdelhameed

Bayoumi

. An efficient deep learning system for epileptic seizure prediction. IEEE Int Symp Circ S. 2021; 1-5.

Jemal

Mezghani

Abou-Abbas

Mitiche

. An interpretable deep learning classifier for epileptic seizure prediction using EEG data. IEEE Access. 2022; 10: 60141-60150.

Gotman

. Automatic detection of seizures and spikes. J Clin Neurophysiol. 1999; 16(2): 130-140.

Tzallas

Tsipouras

Fotiadis

. Epileptic seizure detection in EEGs using time-frequency analysis. IEEE T Inf Technol B. 2009; 13(5): 703-710.

10.

Orhan

Hekim

Ozer

Provaznik

. Epilepsy diagnosis using probability density functions of EEG signals. IEEE INISTA. 2011; 626-630.

11.

Bezobrazova

Golovko

. Comparative analysis of forecasting neural networks in the application for epilepsy detection. IEEE IDAACS. 2007; 202-206.

12.

Entropy

SUP

. Fast and Robust Detection of Epilepsy in Noisy BEG.

13.

Juarez-Guerra

Alarcon-Aquino

Gomez-Gil

. Epilepsy seizure detection in EEG signals using wavelet transforms and neural networks. In New trends in networking, computing, E-learning, systems sciences, and engineering. 2015; 261-269.

14.

Lasefr

Ayyalasomayajula

SSV

Elleithy

. Epilepsy seizure detection using EEG signals. IEEE UEMCON. 2017; 162-167.

15.

Wang

Zhao

Pourpanah

. Recent advances in deep learning. Int J Mach Learn Cyb. 2020; 11: 747-750.

16.

Orhan

Hekim

Ozer

. EEG signals classification using the K-means clustering and a multilayer perceptron neural network model. Expert Syst Appl. 2011; 38(10): 13475-13481.

17.

Korshunova

Kindermans

Degrave

Verhoeven

Brinkmann

Dambre

. Towards improved design and evaluation of epileptic seizure predictors. IEEE T Bio-Med Eng. 2017; 65(3): 502-510.

18.

Nejedly

Kremen

Sladky

Nasseri

Guragain

Klimes

, et al. Deep-learning for seizure forecasting in canines with epilepsy. J Neural Eng. 2019; 16(3): 036031.

19.

Nejedly

Kremen

Sladky

Nasseri

Guragain

Klimes

, et al. Deep-learning for seizure forecasting in canines with epilepsy. J Neural Eng. 2019; 16(3): 036031.

20.

Acharya

Hagiwara

Tan

Adeli

. Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals. Comput Biol Med. 2018; 100: 270-278.

21.

Shoeibi

Moridian

Khodatars

Ghassemi

Jafari

Alizadehsani

, et al. An overview of deep learning techniques for epileptic seizures detection and prediction based on neuroimaging modalities: Methods, challenges, and future works. Comput Biol Med. 2022; 106053.

22.

Hossain

Amin

Alsulaiman

Muhammad

. Applying deep learning for epilepsy seizure detection and brain mapping visualization. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM). 2019; 15(1s): 1-17.

23.

Ibrahim

Djemal

Alsuwailem

Gannouni

. Electroencephalography (EEG)-based epileptic seizure prediction using entropy and K-nearest neighbor (KNN). Comm Sci Tech. 2017; 2(1).

24.

Hussein

Palangi

Ward

Wang

. Optimized deep neural network architecture for robust detection of epileptic seizures using EEG signals. Clin Neurophysiol. 2019; 130(1): 25-37.

25.

Inoue

Nishida

. Deep recurrent neural network for mobile human activity recognition with high throughput. Artificial Life and Robotics. 2018; 23: 173-185.

26.

Dissanayake

Fernando

Denman

Sridharan

Fookes

. Deep learning for patient-independent epileptic seizure prediction using scalp EEG signals. IEEE Sens J. 2021; 21(7): 9377-9388.

27.

Fox

Hightower

Liao

Schulz

Borriello

. Bayesian filtering for location estimation. IEEE Pervas Comput. 2003; 2(3): 24-33.

28.

Chen

Shie

. Fuzzy classification systems based on fuzzy information gain measures. Expert Syst Appl. 2009; 36(3): 4517-4522.

29.

Zhou

Zhao

. Resnext and res2net structures for speaker verification. In 2021 IEEE SLT. 2021; 301-307.

30.

Xie

Girshick

Dollár

. Aggregated residual transformations for deep neural networks. Proc Cvpr IEEE. 2017; 1492-1500.

31.

Wei

Zhao

. Development of a LeNet-5 gas identification CNN structure for electronic noses. Sensors. 2019; 19(1): 217.

32.

CHB-MIT Scalp EEG, “https://physionet.org/content/chbmit/1.0.0/”, accessed on November 2023.