An EEG based real-time epilepsy seizure detection approach using discrete wavelet transform and machine learning methods

https://doi.org/10.1016/j.bspc.2022.103820Get rights and content

Highlights

  • DB4-DWT and DB16-DWT were proposed to extract approximate and details of signals and remove redundant information.

  • Improved the robustness of EEG based epilepsy detection via machine learning methods.

  • Proposed a method that can achieve 97% accuracy and 96.67% sensitivity in 3-class classification (health control, seizure free and seizure active) using Dataset UB, and 96.38% accuracy, 96.15% sensitivity and 3.24% false positive rate in the real-time seizure detection using Dataset CHB-MIT.

  • Implemented an automatic seizure detection approach in real-time way.

Abstract

Epilepsy is one of the most common complex brain disorders which is a chronic non-communicable disease caused by paroxysmal abnormal super-synchronous electrical activity of brain neurons. This paper proposed an electroencephalogram (EEG) based real-time approach to detect epilepsy seizures. Discrete wavelet transform and eight eigenvalues’ algorithms are applied to extract features in different sub-frequency bands. Then support vector machine is employed for three-classes classification of health control, seizure free and seizure active, and finally RUSBoosted tree Ensemble method is used for real-time seizure onset detection. The proposed algorithm is evaluated using two public datasets: one short-term dataset named UB and one long-term dataset named CHB-MIT. The results show that the algorithm achieves 97% accuracy and 96.67% sensitivity in the three-classes classification of health control, seizure-free and seizure-active groups in UB dataset, and 96.38% accuracy, 96.15% sensitivity, 3.24% false positive rate for the real time seizure onset detection in CHB-MIT Dataset.

Introduction

Epilepsy is a chronic non-communicable disease caused by the abnormal synchronous electrical activity of brain neurons [1], [2]. It is also one of the most common neurological diseases in the world, and affects approximately 50 million people [2], [3]. Due to the differences in the starting region and propagation mode of abnormal electrical activity in the brain, the clinical manifestations of epilepsy are diversified and complicated [4]. Repeated seizures can cause persistent negative effects on patients' mental and cognitive functions, and bring life-threatening risks [5]. Therefore, research on the diagnosis and treatment of epilepsy has very important clinical significance. Automatic identification of epilepsy seizures from electroencephalogram (EEG) signals and its real-time implementation can provide an objective reference basis for the diagnosis and in time evaluation of epilepsy, thereby reducing the workload of doctors and improving the efficiency of treatment [6]. Majority of the recent papers have set the ultimate objective of developing automated EEG monitoring system to detect epileptic seizures. Bhattacharyya et al. highlighted a real-time seizure detection through empirical Wavelet transform method [7]. Disruptive EEG networks for epileptic seizures in real-time application reported by Bomela et al. [8]. Harmonic Wavelet packet transform with relevance vector machine method were proposed by Vidyaratne et al. [2].

EEG is a microvolt level electrical signal generated by synchronized neuronal activity in the brain collected by electrodes placed at a specific position on the scalp [9], [10]. EEG abnormalities in epileptic seizures are mainly manifested as spike waves and sharp waves [11]. Therefore, using feature extraction methods to find the eigenvalues which can divide the normal waves and spike or sharp waves in different seizure state should be thoroughly investigated. So far many methods in time, frequency, and time–frequency domains have been developed such as discrete wavelet transform (DWT), empirical mode decomposition (EMD), Q-wavelet transformation, Hilbert-Huang transform (HHT), mean amplitude spectrum (MAS), etc [7], [12], [13], [14], [15]. Another important progress of epilepsy seizure detection is the development of machine learning based classification methods. The main objective of machine learning methods is to overcome the robustness of EEG individuals in epilepsy detection. Specifically, support vector machine (SVM), linear discriminant analysis (LDA), naive Bayes, logistic regression (LR), random forest were used to classify the different seizure states in previous studies [12], [16], [17], [18], [19]. Currently, automatic epilepsy detection can be divided into two types: offline seizure detection and real-time seizure detection. The purpose of offline seizure detection is to identify epileptic seizure signals as accurately as possible from EEG signal [20]. The purpose of real-time seizure detection is to identify seizures onsite with the shortest possible delay when the patient has a seizure during continuous EEG monitoring [21].

Two public datasets are available in EEG seizure detection. Dataset UB is a short-term dataset from the University of Bonn, which is used to do seizure event detection through the classifications of two classes (seizure free and seizure active) and three classes (health control, seizure free and seizure active). In the studies of two-classes classification, fractional linear prediction and SVM were used by Joshi et al. and achieved 95.33% accuracy [22]. Fast Fourier transform with k-nearest neighbor (k-NN) model proposed by Ghaderyan et al. could result in 98.72% accuracy [23]. In addition, The Dual-tree complex wavelet and the nearest neighbor (NN) model was reported to have 95.5% accuracy by Chen et al. [24]. Meanwhile, in the three classes classification studies, Acharya et al. discussed four entropy parameters (approximate entropy, sample entropy, two phase entropies) combined with fuzzy classifier, and achieved 98.1 % accuracy [25]. Omidvar et al. used the DB4-DWT method based on the artificial neural network and SVM models, and got 98.7% accuracy [12]. Currently, feature extraction based on machine learning classification is one of the most researched approaches in seizure event detection using EEG signal. Tunable-Q wavelet transform based multiscale entropy measure proposed by Bhattacharyya, A., et al. is used to classifier 3-classes between normal, seizure-free and seizure EEG signals and achieve 98.6% accuracy results [26]. Gupta, V. and R.B. Pachori stated that Fourier-Bessel series expansion (FBSE) and weighted multiscale Renyi permutation entropy (WMRPE) for EEG rhythms and get 97.3% accuracy results [27]. Empirical wavelet transform (EWT) with FBSE method highlighted by Anuragi, A., et al. can also achieved 97.7% accuracy classification [28].

Dataset CHB-MIT is a long-term dataset from Boston Children’s Hospital, which is used by many researchers to do the real-time automatic seizure detection. Samiee et al. used multivariate textural features with gray-level co-occurrence matrix (GLCM) in SVM and reported 70.19% sensitivity in the real-time seizure detection [29]. As a contrast, time delay embedding method proposed by Zabihi et al. was reported to have 89.01% sensitivity [30]. In particularly, graph theory analysis, function connectivity analysis and effective connectivity analysis were used in the seizure detection [16], [31], [32]. Bomela et al. constructed the network connectivity using Fourier transform to detect the seizure onset and reported 93.6 % sensitivity and a false positive rate of 0.16 per hour (FP/h) result [8]. A stactked 1D-CNN model is presented via Wang, X., et al. to detect seizure onset automatically and achieved 88.14% accuracy and 0.38% false positive (FP) result [33]. Orthogonal matching pursuit with DWT as pre-processing progress with non-linear features and SVM classifier can also detect the seizure onset in same dataset, Zarei, A. and B.M. Asl, used this method reported 96.81% sensitivity and 2.74% FP result [34]. Li, C., et al. proposed EMD, common spatial pattern and SVM model get 97.34% sensitivity, 2.5% FP output as well [35].

In this study, the proposed real time EEG based seizure detection method includes four major steps using the aforementioned both datasets (UB and CHB-MIT) in two experiments. Specifically, in the first experiment using Dataset UB, DWT analysis with DB4 mother wavelet was used to decompose the raw EEG signal data. After feature extraction and selection, 12 eigenvalues were evaluated as the input of the SVM model to classify health control, seizure-free and seizure-active subjects. Based on the first experiment, real-time seizure detection was implemented using Dataset CHB-MIT. Similarly, DB16 DWT analysis with 7 eigenvalues were fed into the SVM and RUSBoosted tree Ensemble model to obtain the final results. All the experiments in this study were carried out in a Dell workstation with dual Intel Xeon E5-2697V3 CPUs using MATLAB 2019b. The main contributions and innovations of this study are: (1) DB4-DWT and DB16-DWT were proposed to extract approximate and details of signals and remove redundant information. (2) Improved the robustness of EEG based epilepsy detection via machine learning methods. (3) Proposed a method that can achieve 97% accuracy and 96.67% sensitivity in 3-class classification (health control, seizure free and seizure active) using Dataset UB, and 96.38% accuracy, 96.15% sensitivity and 3.24% false positive rate in the real-time seizure detection using Dataset CHB-MIT. (4) Implemented an automatic seizure detection approach in real-time way.

The first section of the paper provided a brief introduction of the work. Section II described the details of the short-term dataset (Dataset UB) and long-term dataset (Dataset CHB-MIT). The pre-processing, feature extraction, classification and real-time application are also introduced in this section. Section III reported the work in our experiments and results obtained using the proposed method. Comparisons of previous work using the same datasets were conducted and evaluated in Section IV. Section V concluded the paper.

Section snippets

Methodology

The proposed methodology utilized DWT for the data pre-processing, and calculated nine eigenvalues via entropy-based and statistical measures to extract features. SVM and RUSBoosted tree Enemble methods were used to train and test Dataset UB and Dataset CHB-MIT. The framework of the proposed method is described as follow Fig. 1:

Experiments and results

Accuracy and sensitivity were used to evaluate the 3-class classification for Dataset UB, while accuracy, sensitivity, false positive rate and seizure onset detection delay were used to evaluate the proposed method for Dataset CHB-MIT.

Accuracy is a direct parameter in method evaluation which is define as follow:Acc=TP+TNTP+TN+FP+FNwhere, TP is the true positive, TN is the true negative, FP is the false positive and FN is the false negative.

Sensitivity is another parameter for evaluation which

Discussion

We used DWT to decompose EEG raw signal into different frequency bands, after that nine eigenvalues are calculated in each sub-band. However, not every sub-band’s features are obviously different between seizure active state and seizure free state. Thus, we studied the level of decompositions, mother wavelet selection, and sliding window size selection to get the best performance. The work and results are presented below:

Conclusion

This study proposed an EEG based real-time epilepsy seizure detection approach using DWT, SVM and RUSBoosted tree Ensemble models of machine learning, and evaluated its performance by comparison. Using the 12 eigenvalues in corresponding decomposition levels extracted by DB4-DWT, and SVM models, our study achieved a 97% accuracy and 96.67% sensitivity in the three classes classification (health control, seizure-free and seizure-active) of Dataset UB. Experiments show the proposed method can

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.

References (46)

  • A. Anuragi et al.

    Automated FBSE-EWT based learning framework for detection of epileptic seizures using time-segmented EEG signals

    Comput. Biol. Med.

    (2021)
  • K. Samiee et al.

    Long-term epileptic EEG classification via 2D mapping and textural features

    Expert Syst. Appl.

    (2015)
  • X. Wang et al.

    One dimensional convolutional neural networks for seizure onset detection using long-term scalp and intracranial EEG

    Neurocomputing

    (2021)
  • A. Zarei et al.

    Automatic seizure detection using orthogonal matching pursuit, discrete wavelet transform, and entropy based features of EEG signals

    Comput. Biol. Med.

    (2021)
  • A. Ishikawa et al.

    The fuzzy entropy concept and its application

    Fuzzy Sets Syst.

    (1979)
  • Y. Kumar et al.

    Epileptic seizure detection using DWT based fuzzy approximate entropy and support vector machine

    Neurocomputing

    (2014)
  • A.S. Zandi et al.

    Automated real-time epileptic seizure detection in scalp EEG recordings using an algorithm based on wavelet packet transform

    IEEE Trans. Biomed. Eng.

    (2010)
  • L.S. Vidyaratne et al.

    Real-time epileptic seizure detection using EEG

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2017)
  • D.C. Bergen

    Do seizures harm the brain?

    Epilepsy Curr.

    (2006)
  • R. Rosch et al.

    Network dynamics in the healthy and epileptic developing brain

    Network Neurosci.

    (2018)
  • A. Bhattacharyya et al.

    A multivariate approach for patient-specific EEG seizure detection using empirical wavelet transform

    IEEE Trans. Biomed. Eng.

    (2017)
  • W. Bomela et al.

    Real-time inference and Detection of Disruptive eeG networks for epileptic Seizures

    Sci. Rep.

    (2020)
  • Y.i. Liang et al.

    Altered Functional Connectivity after Epileptic Seizure Revealed by Scalp EEG

    Neural Plasticity

    (2020)
  • Cited by (43)

    • Landscape of epilepsy research: Analysis and future trajectory

      2024, Interdisciplinary Neurosurgery: Advanced Techniques and Case Management
    View all citing articles on Scopus
    1

    0000-0002-6770-4366

    View full text