Automatic detection of epileptic seizure based on approximate entropy, recurrence quantification analysis and convolutional neural networks

https://doi.org/10.1016/j.artmed.2019.101711Get rights and content

Highlights

  • ApEn is used to quantify the complexity of the epileptic seizure.

  • RQA estimates the recurrence behaviors of the epileptic seizure.

  • A new method for automatic epileptic EEG recordings based on the ApEn and RQA combined with CNN is proposed.

  • The classification accuracy using the proposed method can reach to 99.26%.

Abstract

Epilepsy is the most common neurological disorder in humans. Electroencephalogram is a prevalent tool for diagnosing the epileptic seizure activity in clinical, which provides valuable information for understanding the physiological mechanisms behind epileptic disorders. Approximate entropy and recurrence quantification analysis are nonlinear analysis tools to quantify the complexity and recurrence behaviors of non-stationary signals, respectively. Convolutional neural networks are powerful class of models. In this paper, a new method for automatic epileptic electroencephalogram recordings based on the approximate entropy and recurrence quantification analysis combined with a convolutional neural network were proposed. The Bonn dataset was used to assess the proposed approach. The results indicated that the performance of the epileptic seizure detection by approximate entropy and recurrence quantification analysis is good (all of the sensitivities, specificities and accuracies are greater than 80%); especially the sensitivity, specificity and accuracy of the recurrence rate achieved 92.17%, 91.75% and 92.00%. When combines the approximate entropy and recurrence quantification analysis features with convolutional neural networks to automatically differentiate seizure electroencephalogram from normal recordings, the classification result can reach to 98.84%, 99.35% and 99.26%. Thus, this makes automatic detection of epileptic recordings become possible and it would be a valuable tool for the clinical diagnosis and treatment of epilepsy.

Introduction

Epilepsy is a common neurological disorder caused by the abnormal electrical activity in the brain [1]. Recurrent seizures are extremely harmful to the patients’ psychological and mental health. The brain is abnormally excited and over-synchronized when the seizures occur, and epilepsy causes temporary malfunction of central nervous system. Electroencephalogram (EEG) is recordings of spontaneous bio-potential of the brain, which could catch useful information for detecting epilepsy. Therefore, EEG has been widely used in diagnosis and treatment of epilepsy [[2], [3], [4]].

Usually the medical experts adopt the method of artificial interpretation to detection the seizure. However, visual inspection by doctors for distinguishing EEGs is a particularly time-consuming process, and it is too easily influenced by the level of doctors and subjective factors. Thus, it is necessary to develop automatic seizure recognition approaches. Previous researchers had investigated this problem, such as Ouyang et al. [5] use recurrence plot metrics to determine epilepsy rats and there achieve a high accuracy, Ma et al. [6] used sample entropy and artificial neural network to predict seizure. However, no one discuss the result of combined nonlinear analysis methods with convolutional neural networks.

Kinds of evidences indicate that the brain is a complex non-linear dynamical system, and the EEG signals are non-linear and non-stationary [[7], [8], [9], [10], [11]]. Recurrence plot is an nonlinear tool to analyze the recurrence characteristics of time series in phase space [12,13]. Recurrence plot performs much better than traditional methods for the analysis of nonlinear dynamical systems, especially for short-time series. In recent years, recurrence quantitative analysis (RQA) has been widely used to analyze physiological signals, such as EEG recordings [[14], [15], [16], [17]], electromyography signals [18], heart rate variability [19], blood pressure [20] etc. Acharya et al. [17] analyzed the EEG signals during sleep with RQA and found that the non-linear characteristics of EEG signals in different sleep stages showed significant difference. Billeci et al. [21] and Yang et al. [22] reported that it is feasible to predict seizure in advance by RQA. Another nonlinear metric is Approximate entropy (ApEn) [23], which is commonly used to measure the regularity of time series. ApEn has been widely applied to evaluate the complexity of epileptic seizure time series [24]. The value of the ApEn decreased abruptly during an epileptic activity as a result of the synchronous discharge of a large number of neurons [25].

It is important to choose the features that can represent the characteristics of the EEG recording for detecting seizures. Hence, in this work, RQA and ApEn are used to discriminate the EEGs.

Convolutional neural networks (CNN) is one of the most common classifiers [26,27], especially for classifying EEGs [[28], [29], [30]], which is chosen as the classifier model for this study. CNN can be defined as highly inter-connected structure comprising of adaptive simple processing elements which are called neurons or nodes [31,32]. Besides, CNN are computational modeling tools that could compute data processing and knowledge representation in parallel on a massive scale. CNN consists of many computational neural units connected to each other. In order to generate the desired mapping, the CNN has to be trained to correct the biases. Recently, CNN is widely used in data analysis and diagnose in the biomedical field [[33], [34], [35], [36], [37]]. Especially, the CNN is suitable for pattern recognition problems based on the ability of learning from examples, and reproduce any non-linear functions of input and the highly parallel and regular structure.

The training algorithm is an important component of the CNN. If training with impertinent algorithm, a good topology maybe preforms inefficient. In current work, Bayesian regularization back-propagation is used as the training algorithm [38,39], which updates the weight and bias values according to Levenberg-Marquardt optimization. It takes advantage of a combination of minimized squared errors and weights, and then selects the correct combination to produce a generalized network.

In this paper, an epilepsy EEG classification method based on recurrence quantization analysis, approximate entropy and convolutional neural networks is proposed. First, ApEn and RQA are used to classify the non-seizure group form seizure subjects. Then use the combination of ApEn values and the RQA quantified values as the input of CNN to realizes the automatic detection of epileptic EEG. In the feature extraction part, extracting RQA quantitative values and ApEn values are used as nonlinear features. In the clinical EEG data experiment, the model has a great performance.

Section snippets

Materials

The experimental data was obtained from the EEG database of the Bonn Epilepsy Laboratory in Germany [40], which is consist of five sets (named as Z, O, N, F and S). Each set contains 100 clinical intracranial EEG recordings of 23.6 s duration, with sampling frequency of 173.6 Hz. After visual inspection for artifacts, select these segments and cut from successive multi-channel EEG recordings. Surface EEG recordings that were carried out on five healthy volunteers by a standardized electrode

Analysis by ApEn

Each sample was processed by ApEn, and the values were directly extracted for automatic classification of epileptic EEG. To further analyze the differences in ApEn between epileptic EEG and normal EEG signals, the ApEn distribution maps and box plots were summarize in Fig. 3.

The value of ApEn of EEG recordings from healthy subjects show significantly greater than those in seizures (p < 0.0001).

Results of RQA

From Fig. 4 to Fig. 9, the RQA eigenvalues of EEG in seizures were significantly greater than those in

Discussions

At present, the detection of epilepsy is mainly done by doctors through visual inspection based on EEG. Due to visual inspection is time-consuming and inefficient, automatic detection of epileptic EEG recordings are of great clinical significance which could reduce the workload of medical professionals. Considering that the brain is a very complex nonlinear system, use nonlinear measures to investigate the EEG signals could better reflect the intrinsic nature of the system. This article

Conclusions

In this paper, the ability of two nonlinear indices, that is ApEn and RQA, combination with CNN to detect epileptic seizure in EEGs is investigated. The EEG recordings are first analyzed by the two nonlinear metrics. Then, use the nonlinear results as inputs and use CNN as classifier for classifying the normal EEG recordings from seizure EEGs. The best classification accuracies are obtained by this novel process. This makes detection of seizure in clinical epilepsy diagnostics in real-time

Declaration of Competing Interest

The authors declare no competing interests.

References (56)

  • U.R. Acharya

    Automated EEG-based screening of depression using deep convolutional neural network

    Comput Methods Programs Biomed

    (2018)
  • J. Gu

    Recent advances in convolutional neural networks

    Pattern Recognit

    (2018)
  • A. Qayyum

    Medical image retrieval using deep convolutional neural network

    Neurocomputing

    (2017)
  • M. Vardhana

    Convolutional neural network for bio-medical image segmentation with hardware acceleration

    Cogn Syst Res

    (2018)
  • M.W. Gardner et al.

    Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences

    Atmos Environ

    (1998)
  • N. Marwan

    Recurrence plots for the analysis of complex systems

    Phys Rep

    (2007)
  • M. Casdagli

    Recurrence plots revisited

    Physica D

    (1997)
  • L. Guo et al.

    Epileptic seizure detection using multiwavelet transform based approximate entropy and artificial neural networks

    J Neurosci Methods

    (2010)
  • E.D. Übeyli

    Combined neural network model employing wavelet coefficients for EEG signals classification

    Digit Signal Proc

    (2009)
  • P. van Mierlo

    Network perspectives on epilepsy using EEG/MEG source connectivity

    Front Neurol

    (2019)
  • Gaoxiang Ouyang et al.

    Using recurrence plot for determinism analysis of EEG recordings in genetic absence epilepsy rats

    Clin Neurophysiol

    (2008)
  • Ma Li et al.

    A preliminary study on epileptic seizure prediction using sample entropy and artificial neural network

    Chin J Biomed Eng

    (2013)
  • D. Garrett

    Comparison of linear, nonlinear, and feature selection methods for EEG signal classification

    Ieee Trans Neural Syst Rehabil Eng

    (2003)
  • K. Natarajan

    Nonlinear analysis of EEG signals at different mental states

    Biomed Eng Online

    (2004)
  • J. Eckmann et al.

    Recurrence plots of dynamical systems

    World Scient Ser Nonlinear Sci Ser A

    (1995)
  • L. Priano

    Sleep microstructure analysis using Recurrence Plots: evidence for a subcortical oscillator?

    Europhys Lett

    (1987)
  • U.R. ACHARYA

    Application of recurrence quantification analysis for the automated identification of epileptic eeg signals

    Int J Neural Syst

    (2011)
  • U.R. Acharya

    Non-linear analysis of EEG signals at various sleep stages

    Comput Methods Programs Biomed

    (2005)
  • Cited by (0)

    This paper belongs to Special Issue: Deep learning Methods for Medical Applications.

    View full text