Elsevier

Future Generation Computer Systems

Volume 127, February 2022, Pages 426-434
Future Generation Computer Systems

Real-time epileptic seizure recognition using Bayesian genetic whale optimizer and adaptive machine learning

https://doi.org/10.1016/j.future.2021.09.032Get rights and content

Highlights

  • New model is proposed to recognize seizure from time series in the cloud framework.

  • NB-GWOA proposes to select the optimal features from the epileptic seizure dataset.

  • GA uses to enhance the exploitation of WOA, while NB uses to assess every agent.

  • ELM guides by DE to provide optimal parameters for an efficient leaning solution.

  • This model assesses by different evaluation criteria and outcomes recent DL methods.

Abstract

The electroencephalogram (EEG) has been commonly used to identify epileptic seizures, but identification of seizures from EEG remains a challenging task that requires qualified neurophysiologists. It is important to detect seizures in real time, which can be achieved in an internet of things (IoT)-based cloud platform to alert patients of impending seizures. Therefore, in this study, we propose a new model to recognize seizure states (e.g., ictal, preictal, interictal) from EEG in the IoT framework to monitor patients remotely. The proposed model uses an efficient hybrid genetic whale optimization algorithm (GWOA) based on naïve Bayes (NB-GWOA) for feature selection, and an adaptive extreme learning machine (ELM) based on a differential evolutionary (DE) algorithm (DEELM) for classification. In the NB-GWOA method, the genetic algorithm serves to enhance the exploitation of the whale optimization algorithm in the search of the optimal solutions, while the naïve Bayes method is used to determine a fitness function to assess every agent in the search space. GWOA has strong robustness and is capable of finding the best solutions in less than five iterations, so it is suitable for selecting discriminative features from a huge number of neurofeatures obtained from EEG. Further, the classification model is constructed based on ELM, which uses the DE algorithm for a fast and efficient learning solution. Results show that the proposed NB-GWOA-DEELM model can avoid over- and under-fitting and can provide better and more accurate performance in classifying seizure states from EEG than its competitors.

Introduction

Epilepsy is a chronic neurological disorder that is characterized by recurrent seizures. Understanding and identifying epileptic seizures is a difficult task in the field of mental health. A seizure consists of abnormal brain activity, which may result in a loss of consciousness or convulsions [1]. The most useful tool for epilepsy detection is the electroencephalogram (EEG), which directly records the brain’s electrical activity [2]. Identification of epilepsy using EEG signals is a challenging task requiring highly trained neurophysiologists. Very few patients with epilepsy respond to anticonvulsant medication. As a result, individuals with epilepsy have no way of knowing when seizures will occur while they are engaging in any other activity that would make seizures particularly dangerous [3]. Thus, seizure recognition systems may be able to help patients with epilepsy to lead normal lives. To make an EEG-based seizure recognition system works effectively, computational algorithms and machine learning (ML) techniques must reliably identify periods of the increased likelihood of seizure occurring. If these seizure-permissive brain states can be identified, they can be uploaded to cloud computing in a smart city to alert patients of impending seizures. As a result, patients could avoid potentially dangerous behavior, and medications could be administered whenever they are needed to prevent impending seizures, reducing overall side effects. Identification of seizures is therefore of significant importance for improving the quality of life of patients, as early warning and early detection can result in timely treatment [4].

EEG signals provide valuable information about different physiological states of the brain, and are useful for understanding brain activity and dysfunction. The indication of nonlinear deterministic and finite-dimensional structures in EEG is a difficult and complicated issue. There are four seizure-based states: interictal (between seizures), preictal (before seizures), ictal (seizure), and postictal (after seizures) [5], [6]. EEG-based recognition of these seizure-based states is important in clinical practice. In particular, the ability to accurately identify a preictal state apart from interictal, ictal, and postictal states is required for seizure forecasting.

Existing methods in literature use various classifiers and EEG features to increase the precision of epilepsy identification, but these methods are still limited in some perspectives, such as time-consuming for parameters tuning and adaption, sensitive to noise, stuck in local minima specially with emergency cases, and only designed for binary classification. In addition, the parameters of existing methods should be optimized to find the optimal values to fit the new data. The optimized parameters enable the model to perform continuously in real time application. Hence, most existing methods are not suitable for epileptic seizure recognition in an IoT system. Many of published results are based on binary classification (normal vs. epileptic). For example, [7] distinguished epileptic seizure from eyes-open resting state using the classical decision tree classifier. Some other works are based on three-class classification (normal, pre-ictal, and epileptic). For example, in [8], [9], [10], [11], [12], this research group has reported and compared a series of classification methods for three-class classification. Taking [10] for example, the authors compared seven different classical classifiers: Fuzzy Sugeno Classifier, support vector machine, k-nearest neighbor, probabilistic neural network, decision tree, Gaussian mixture model, and naïve Bayes with different features, and they found that Fuzzy Sugeno classifier achieved the best results. However, the traditional models rely on a local-side standalone application, and based on generic rules, are often tailored for a particular case [13], [14], [15], [16]. Such models do not accommodate a large amount of data when monitoring patients with epilepsy. Hence, an alternative model which can offer an efficient and accurate solution for controlling seizures is required. Such a model could be integrated into the design of an epileptic monitoring platform, providing architecture and description of the seizure assessment prototype based on the internet of things (IoT) smart healthcare model [15], [17]. From most of previous studies our work is different because it was aimed to predict ictal states of epileptic seizure in an IoT framework to warn users beforehand or to notify doctors in a nearly real-time manner. As a consequence, the new model should be executed in a fast and accurate manner. Also, the new model is supposed to work continuously for a long time so it should be able to classify all possible states rather than to distinguish ictal states from normal states (a binary classification)

In existing IoT-based seizure monitoring systems, ML techniques play an important role [18], [19], [20]. Particularly, feature selection is an essential and crucial task for building an efficient model for epileptic prediction because EEG signals produce high dimensional, highly correlated features in vast numbers, which could greatly decrease the detection accuracy [21]. Thus, it is desirable to develop a feature selection method to identify the optimal number of features to ensure the continuity of the model’s work in real time with the highest efficiency and speed. Some feature selection methods have been proposed using a global search-based swarm intelligence (SI) for seizure detection [22], [23]. Because of the high performance of SI methods in handling high-dimensional and complex data, SI algorithms have been successfully applied in many IoT applications [24], [25], [26], [27]. However, there are few studies concerning the use of SI algorithms for epileptic monitoring. One important reason is that most SI algorithms have problems such as high computational time, getting stuck in local minima, failing to converge to the optimal solution, and sensitivity to uncertainty and missing data [28].

The aim of this study is to develop a new model to monitor epileptic patients in a real-time manner. To solve the problem of feature selection for an IoT-based cloud monitoring system, we need a new optimal feature selection method, which should not only have high accuracy but should also have high convergence speed and low computational complexity. To this end, we propose a new model based on the whale optimization algorithm (WOA) [29] and the genetic algorithm (GA) [30], which can perform global and local searches, respectively, to efficiently capture distinctive information for distinguishing EEG brain signals. This new method combining the GA and the WOA is named genetic whale optimization algorithm (GWOA). The basic ideas of the WOA and the GA are briefly introduced as follows. The WOA simulates humpback whales’ living behavior, and it has an excellent performance in terms of improved exploration, avoidance of local optima, and high convergence speed. Literature shows the competitiveness of the WOA for optimization problems and engineering applications [31], [32], [33], [34]. In addition, WOA shows good performance in different applications of EEG signals classification such as [35], [36]. In [35], the WOA is used to compensate for the deficiency of the random weight initialization for the basic ELM. The WOA-ELM method was used in their work for EEG classification of BCI, and a high accuracy was achieved. In [36], proposed the WOA-SVM model for binary epileptic EEG classification (normal vs. abnormal) and WOA is used to optimize the parameters of SVM as well as to select features. This WOA-SVM model had two main phases: in the first phase, WOA was used to estimate the optimal subset of features, and in the second phase, the parameters of the RBF kernel function in SVM classifier was optimized by WOA.

However, the WOA is somewhat limited in its capability of exploitation and local search. To improve the performance of the WOA, the GA, which is an evolutionary algorithm to solve optimization problems, is incorporated into the WOA for local search. GAs rely on randomized operations and use probabilistic selection rules, not deterministic ones [30]. The GA is embedded within the WOA as an operator to search for the best solution at the next iteration in the neighbor of both the randomly selected solution and the best solution at the current iteration to increase the exploitation. As a result, the proposed GWOA uses a WOA and a GA for global and local searches, respectively, to achieve a better optimization performance. Furthermore, the naïve Bayes (NB) criterion is used as a part of the GWOA optimizer to assess every agent in the search space and to evaluate the appropriateness of the selected neurofeatures [23]. The NB-based fitness function is used for both the WOA and the GA as a part of the optimizer to evaluate each agent’s position in the search space. The resultant NB-GWOA model is expected to increase the performance (in terms of accuracy and speed) of the epilepsy detection from EEG brain signals.

Furthermore, we need a fast and precise classification model to recognize epileptic states from selected EEG features. The extreme learning machine (ELM) method is commonly used as a feature-based classification system for various signal processing and pattern recognition problems [34], [37]. The ELM is described as a single hidden layer neural network with randomly initialized hidden neuron weights that possess universal approximation capability. For simple applications, the ELM is more effective and computationally faster than other traditional ML methods, and it has a better robustness property and a faster training speed than deep learning methods [38], [39], [40]. However, the main drawback of the ELM is its random nature that may provide unpredictable results, take a long time to find optimal parameters, and become stuck in local minima, especially for complicated applications. Therefore, the differential evolution (DE) algorithm is proposed to improve the capability of the ELM by optimizing hidden-layer parameters (weights and biases) [41]. This new DEELM method has the potential to build a powerful detection model in a real-time process for epileptic seizure monitoring.

In this study, this proposed NB-GWOA-DEELM model can detect the epileptic seizure status of a patient and recognize different states (awake states, preictal states, interictal states, and ictal states) in a real-time manner. The GWOA inherits the advantages of the WOA so that it can provide smaller errors in training, achieve high accuracy in fewer iterations, and avoid local minima. Further, the ELM is guided by the evolutionary DE algorithm for optimal hyper-parameters (weights and biases). Our results prove that the NB-GWOA-DEELM model is more robust, has a high convergence speed to an optimal solution, provides fewer errors in training, reduces time consumption, and achieves higher precision in fewer iterations than other ML methods. Therefore, the NB-GWOA-DEELM model is effective in detecting preictal state and has the potential to be used for IoT-based epilepsy monitoring. As shown in Fig. 1, the proposed NB-GWOA-DEELM model could serve as a central data analysis part in an IoT-based cloud computing system for epileptic seizure monitoring, which mainly comprises the patient in smart healthcare, the cloud, and the doctors.

The remaining structure of this paper is organized as follows: The preliminary concepts and methods involved are introduced and the proposed NB-GWOA and DEELM are introduced in Section 2. Section 3 introduces the pipeline of the proposed NB-GWOA-DEELM model for epilepsy detection. Section 4 presents the results of the proposed model and other models under comparison. Finally, our conclusions and proposed future work are presented in Section 5.

Section snippets

Materials and methods

In this section, the EEG dataset is described, and the mathematical models of the WOA, GA, and ELM are introduced first. Next, based on the WOA, GA, and ELM, the proposed NB-GWOA and DEELM models are introduced.

NB-GWOA-DEELM for seizure recognition

The architecture of the proposed NB-GWOA-DEELM model to detect the state of seizure is shown in Fig. 4, and the steps of this model are shown in Algorithm 1. As depicted in Fig. 4, the proposed NB-GWOA-DEELM model consists of five main phases.

In the first phase, the EEG signals are initially pre-processed using routine steps (sampling, segmenting, removing artifact, low-pass filter, and normalization). Then the problem of the imbalanced data is treated using the SMOTE method.

In the second phase

Parameters and computer setting

The proposed model was developed and tested using MATLAB R2016a, on an Intel ®-Core-TM-i7, 2.5 GHz processor, with 8 GB memory. Table 1 presents the description of different parameters initialization for the proposed NB-GWOA-DEELM model.

Feature selection based on NB-GWOA

Table 2 shows the results obtained using the NB-GWOA as compared with conventional WOA based on NB. The embedding of a GA in the WOA improved the accuracy results in terms of worst fitness, best fitness, mean fitness, standard derivation, t-test value (Pttest),

Conclusion and future work

In this study, we propose a new NB-GWOA-DEELM model for the detection of seizures in patients based on their EEG. The experimental results show that the proposed model is highly effective in understanding complex EEG neurodynamic signals, which leads to improved accuracy of epilepsy identification under various measurement criteria compared to existing conventional machine learning and deep learning algorithms. The superiority of the NB-GWOA-DEELM algorithm makes it possible to improve the

CRediT authorship contribution statement

Ahmed M. Anter: Funding acquisition, Formal analysis, Interpretation of data, Conceptualization, Methodology, Creation of model, Software, Writing – original draft. Mohamed Abd Elaziz: Writing – review & editing, Investigation, Validation, Verify the experimental results. Zhiguo Zhang: Conceptualization, Design of study, Formal analysis, Visualization, Writing – review & editing, Investigation, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This research was supported in part by National Natural Science Foundation of China under Grant 81871443, in part by the Science, Technology and Innovation Commission of Shenzhen Municipality Technology Fund under Grant 2021SHIBS003 and JCYJ20170818093322718, in part by Shenzhen Peacock Plan under Grant KQTD2016053112051497. Also, we thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of this manuscript.

Ahmed M. Anter received M.S. and Ph.D. degrees in computer science from Mansoura University in 2010 and 2016, respectively. He is worked in CITC, Mansoura University as a team leader software development from 2006–2010, then he worked in Faculty of Computer Science and Information System, Jazan University, Saudi Arabia as a lecturer from 2011–2014. He is currently an Associate Professor with the Faculty of Computers and Artificial Intelligence, Beni-suef University, Benisuef, Egypt from

References (47)

  • AnterA.M. et al.

    Multi-target QSAR modelling of chemo-genomic data analysis based on Extreme Learning Machine

    Knowl.-Based Syst.

    (2020)
  • WangJ. et al.

    A novel hybrid system based on a new proposed algorithm-Multi-Objective Whale Optimization Algorithm for wind speed forecasting

    Appl. Energy

    (2017)
  • KaurG. et al.

    Chaotic whale optimization algorithm

    J. Comput. Des. Eng.

    (2018)
  • YangP. et al.

    Ensemble of kernel extreme learning machine based random forest classifiers for automatic heartbeat classification

    Biomed. Signal Process. Control

    (2021)
  • RongH.J. et al.

    A fast pruned-extreme learning machine for classification problem

    Neurocomputing

    (2008)
  • WangG. et al.

    Detection of epilepsy with electroencephalogram using rule-based classifiers

    Neurocomputing

    (2017)
  • VandecasteeleK. et al.

    Automated epileptic seizure detection based on wearable ECG and PPG in a hospital environment

    Sensors

    (2017)
  • StafstromC.E. et al.

    Seizures and epilepsy: an overview for neuroscientists

    Cold Spring Harbor Perspect. Med.

    (2015)
  • StamoulisC. et al.

    Noninvasive seizure localization with single-photon emission computed tomography is impacted by preictal/early ictal network dynamics

    IEEE Trans. Biomed. Eng.

    (2018)
  • FaustO. et al.

    Automatic identification of epileptic and background EEG signals using frequency domain parameters

    Int. J. Neural Syst.

    (2010)
  • AcharyaU.R. et al.

    Automatic detection of epileptic EEG signals using higher order cumulant features

    Int. J. Neural Syst.

    (2011)
  • AcharyaU.R. et al.

    Automated diagnosis of epileptic electroencephalogram using independent component analysis and discrete wavelet transform for different electroencephalogram durations

    Proc. Inst. Mech. Eng. H

    (2012)
  • AcharyaU.R. et al.

    Automated diagnosis of epileptic electroencephalogram using independent component analysis and discrete wavelet transform for different electroencephalogram durations

    Proc. Inst. Mech. Eng. H

    (2013)
  • Cited by (31)

    • Success history based adaptive multi-objective differential evolution variants with an interval scheme for solving simultaneous topology, shape and sizing truss reliability optimisation

      2022, Knowledge-Based Systems
      Citation Excerpt :

      For multi-objective optimisation, well-known algorithms include Non-dominated Sorting Genetic Algorithm II (NSGA-II) [31], Pareto Archived Evolution Strategy (PAES) [32], Differential Evolution for Multi-objective Optimisation (DEMO) [33], Multi-Objective Evolutionary Algorithm based Decomposition (MOEA/D) [34], Pareto Envelope based Selection-II (PESA-II) [35], Strength Pareto Evolutionary Algorithm 2 (SPEA2) [36], Unrestricted Population-Size Evolutionary Multi-objective Optimisation algorithm (UPS-EMOA) [37], Differential Evolution for Multi-objective Optimisation based on Gaussian Process models (GP-DEMO) [38], multi-objective multi-verse optimisation algorithm (MOMVO) [39], multi-objective grasshopper optimisation algorithm (MOGOA) [40], multi-objective dragonfly optimisation (MODA) [41], multi-objective salp swarm algorithm (MSSA) [42], hybrid real-code population-based incremental learning and differential evolution algorithm (RPBILDE) [43,44], multi-objective meta-heuristic with iterative parameter distribution estimation (MMIPDE) [45], Success History-based Adaptive Multi-Objective Differential Evolution (SHAMODE) and its hybridised version with the Whale Optimisation Algorithm (SHAMODE-WO) [6]. There are a great number of investigations on improving the efficiency of metaheuristics by adaptive strategies, external archives, and hybridisation between algorithms [6,46–71]. The reliability approximation for using HHO is presented in Zhong et al.

    View all citing articles on Scopus

    Ahmed M. Anter received M.S. and Ph.D. degrees in computer science from Mansoura University in 2010 and 2016, respectively. He is worked in CITC, Mansoura University as a team leader software development from 2006–2010, then he worked in Faculty of Computer Science and Information System, Jazan University, Saudi Arabia as a lecturer from 2011–2014. He is currently an Associate Professor with the Faculty of Computers and Artificial Intelligence, Beni-suef University, Benisuef, Egypt from 2015–now. He has over 35 scientific research publications and serves as a reviewer for various international journals and conferences. His main research interests include pattern recognition and intelligent systems, machine learning, meta-heuristics, and optimization, fuzzy systems. Finally, Anter joined Shenzhen University as post-doctor fellow with the School of Biomedical Engineering, China.

    Mohamed Abd Elaziz received the B.S. and M.S. degrees in computer science and the Ph.D. degree in mathematics and computer science from Zagazig University, Egypt, in 2008, 2011, and 2014, respectively. From 2008 to 2011, he was an Assistant Lecturer with the Department of Computer Science. He is currently an Associate Professor with Zagazig University. He has authored or coauthored more than 100 articles. His research interests include metaheuristic technique, cloud computing machine learning, signal processing, image processing, and evolutionary algorithms.

    Professor Zhang Zhiguo is with School of Biomedical Engineering, Health Science Center, Shenzhen University. He received his Ph.D. degree from the Department of Electrical and Electronic Engineering, The University of Hong Kong, in 2008. He worked as an Assistant Professor with Nanyang Technological University, Singapore and a Professor with Sun Yat-sen University, China. His research interests include brain signal and image analysis, neuromodulation, and brain-inspired computation and his research is now focused on the diagnosis and intervention of brain disorders using advanced data analytical tools. He has published more than 90 journal papers and more than 100 conference papers.

    View full text