Elsevier

Computers in Biology and Medicine

Volume 100, 1 September 2018, Pages 270-278
Computers in Biology and Medicine

Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals

https://doi.org/10.1016/j.compbiomed.2017.09.017Get rights and content

Highlights

  • Classification of normal, preictal, and seizure EEG signals.

  • Performed 13-layer deep convolutional neural network.

  • Implemented ten-fold cross-validation strategy.

  • Obtained accuracy of 88.7%, sensitivity of 95% and specificity of 90%.

Abstract

An encephalogram (EEG) is a commonly used ancillary test to aide in the diagnosis of epilepsy. The EEG signal contains information about the electrical activity of the brain. Traditionally, neurologists employ direct visual inspection to identify epileptiform abnormalities. This technique can be time-consuming, limited by technical artifact, provides variable results secondary to reader expertise level, and is limited in identifying abnormalities. Therefore, it is essential to develop a computer-aided diagnosis (CAD) system to automatically distinguish the class of these EEG signals using machine learning techniques. This is the first study to employ the convolutional neural network (CNN) for analysis of EEG signals. In this work, a 13-layer deep convolutional neural network (CNN) algorithm is implemented to detect normal, preictal, and seizure classes. The proposed technique achieved an accuracy, specificity, and sensitivity of 88.67%, 90.00% and 95.00%, respectively.

Introduction

According to the World Health Organization (WHO), nearly 50 million people suffer from epilepsy worldwide [1]. It is estimated that 2.4 million people are diagnosed with epilepsy annually [1].

Seizures are due to the uncontrolled electrical discharges in a group of neurons [2], [3]. The excessive electrical discharges result in the disruption of brain function. Epilepsy is diagnosed when there is recurrence of at least two unprovoked seizures. It can affect anyone at any age [4].

A timely and accurate diagnosis of epilepsy is essential for patients in order to initiate anti-epileptic drug therapy and subsequently reduce the risk of future seizures and seizure-related complications [5]. Currently, the diagnosis of epilepsy is made by obtaining a detailed history, performing a neurological exam, and ancillary testing such as neuro-imaging and EEG. The EEG signals can identify inter-ictal (between seizures) and ictal (during seizure) epileptiform abnormalities.

Fig. 1 shows a graphical representation of the electrical activity in the brain of healthy subjects and seizure patients. Typically, neurons communicate through electrical signals. Therefore, in a regular brain activity, these electrical signals are normally regulated [3] (see the normal activity in Fig. 1). However, during seizure, there is an abnormally increased hyper-synchronous electrical activity of epileptogenic neural network. This activity may remain localized to one part of the brain, or spread to the entire brain. In either scenario, an individual may experience a clinical seizure (see the seizure activity in Fig. 1) [3]. Neurologists scrutinize the EEG via direct visual inspection to investigate for epileptiform abnormalities that may provide valuable information on the type and etiology of a patient's epilepsy.

However, interpretation of the EEG signals by visual assessment is time-consuming particularly with the increased use of out-patient ambulatory EEG's and in-patient continuous video EEG recordings, where there are hours or days worth of EEG data that needs to be reviewed manually [6]. The majority of EEG software includes some form of automated seizure-detection, however, due to the poor sensitivity and specificity of the pre-determined seizure detection algorithms, the current forms of automated seizure detection are rarely used in clinical practice. In addition, the inherent nature of visual inspection results in varying clinical interpretations based on the EEG reader's level of expertise in electroencephalography. Complicating matters, the quality of the study may be confounded by interfering artifactual signal limiting the reader's ability to accurately identify abnormalities. Moreover, the low yield of routine out-patient studies poses another problem. A patient with epilepsy can go for an outpatient EEG and the study may be completely normal. This is because the brains of patients with epilepsy are generally not continually firing off epileptic discharges. An EEG is simply a “snapshot” of their brain at the moment of recording. The sensitivity of identifying epileptic discharges can be increased by having the patient come back for repeated outpatient studies or recording them for longer periods of time, either via a home ambulatory study or an inpatient continuous video EEG monitoring study, which are both costly and time-intensive for the patient and for the physician reading the EEG.

Patients are referred to an epilepsy monitoring unit for inpatient continuous video EEG monitoring for a couple different reasons. Usually it is done when the diagnosis of epilepsy is not clear, i.e., a patient's history is atypical for a seizure disorder or could represent another condition clinically similar to seizure, i.e. syncope, or if there is no improvement in seizure frequency following anti-epileptic drug administration. The patient is admitted to the hospital and hooked up to an EEG for several days. If they are on anti-epileptic medications, they are discontinued. The whole point is for them to have their seizure or seizure-like event while they are hooked up to the EEG machine. Then, a trained neurologist or epileptologist analyzes the clinical characteristics of the event in conjunction with visual inspection of the EEG to determine if the patient has epilepsy or not. The increased amount of data also allows the epileptologist to look for inter-ictal abnormalities. Generally, during a seizure, the EEG activity becomes very abnormal and can be clearly determined whether the patient's events are epileptic or non-epileptic. The epileptic EEG signals are more chaotic and varies more as compared to the normal EEG signals. During seizure, there is a sudden surge in neural discharge resulting in the increase of disparities in EEG signals. The neurons in the cerebral hemispheres during a seizure misfire and produce abnormal electrical activity. Thus, the number of neurons available for data processing during seizures decreases. However, at times this can be very difficult because some patients with epilepsy, usually patients with frontal lobe epilepsy or seizures emanating from a deep source or a very small area, can have a clinical seizure and the ictal EEG is normal. Although epileptiform abnormalities are invariably present, these abnormalities do not register on scalp surface electrodes. This can therefore make the diagnosis very difficult. In patients with intractable epilepsy undergoing surgical evaluation, invasive intracranial depth electrodes and/or grids are utilized to identify the epileptogenic zone, which carries peri-procedural risks and complications.

Fig. 2 displays sample normal, interictal, and seizure EEG signals from the Bonn University database. The visual interpretation of these signals is prone to inter-observer variabilities. Therefore, for an accurate, fast, and objective diagnosis a computer-aided diagnosis (CAD) system is advocated. Since the seminal article by Adeli et al. [7], automated EEG-based seizure detection and epilepsy diagnosis has been the subject of significant research. Many researchers have proposed different approaches to automatically detect epileptic seizure using EEG signals. For reviews of this literature, see Acharya et al. [8] and Faust et al. [9] where different approaches, namely, time, frequency, time-frequency, and nonlinear methods are discussed. Acharya et al. [10] also review application of entropies for automated EEG-based diagnosis of epilepsy.

The EEG signal is nonlinear and nonstationary in nature thus; the signal is highly complex and is difficult to visually interpret the signals (see Fig. 2). Based on the reviews [8], [9], [10], it can be observed that the researchers have extracted features, performed statistical analysis, ranked the features, and classified the best classifier by comparing the performance of different classifiers. Thus, the workflow consists of many standard steps. This work proposes implementation of deep learning for an automated detection of normal, preictal, and seizure EEG signals without performing feature extraction and selection.

Deep learning is a machine learning technique based on representation learning where the system automatically learns and discovers the features needed for classification from the processing of multiple layers of input data [11]. Deep learning has already proven its capability and has outperformed humans in audio and image recognition tasks [11], [12]. It has been used in many other diverse complicated machine learning applications such as early diagnosis of the Alzheimer's disease [13], prediction of sale prices of real estate units [14], estimation of concrete compressive strength [15]. Moreover, many large technology companies such as Apple, Baidu, Google, IBM, Facebook, Microsoft, and Netflix have embraced and utilized deep learning in their research [16], [17], [18]. In this study, a deep learning method is employed to automatically identify the three classes of EEG signals. To the best the authors' knowledge, this is the first EEG study to employ deep learning algorithm for the automated classification of three EEG classes. A 13-layer deep convolutional neural network (CNN) is developed to categorize the normal, preictal, and seizure class.

Section snippets

Data

EEG segments used in this research are those collected by Andrzejak et al. [19] at Bonn University, Germany (http://epilepsy.uni-freiburg.de/database). The segments were selected from continuous multichannel EEG recordings with artifacts removed via visual examination due to muscle activity and eye movements.

The dataset obtained from 5 patients contains three classes of data, namely, normal (Set B), preictal (Set D), and seizure (Set E). There is a total of 100 EEG signals in each dataset. Each

Pre-processing

Each EEG signal is normalized with Z-score normalization, zero mean and standard deviation of 1 before feeding into the 1-D deep convolutional network (CNN) for training and testing. The sampling rate of the EEG signal is set at 173.61 Hz.

Artificial neural network (ANN)

Generally, an ANN has three layers: input, hidden, and output layers (see Fig. 3) [20]. The concept of ANN is inspired by complex networks structure found in human brains. ANN is made up of a collection of connected units called nodes or neurons. Just like the

Results

The proposed algorithm was implemented on a workstation with two Intel Xeon 2.40 GHz (E5620) processor and a 24 GB random-access memory (RAM) using the MATLAB programming software. It typically took about 12.8 s to complete an epoch of training.

The tabulated confusion matrix across all ten-folds is presented in Table 2. It is observed that 90% of the normal EEG signals are correctly classified as normal EEG signals. Further, a small percentage of 1% and 9% of the normal EEG signals are wrongly

Discussion

Table 4 presents a summary of studies conducted in the automated detection of normal, preictal, and seizure EEG signals obtained from the Bonn University database.

Adeli et al. [34] presented a wavelet-chaos approach to analyze 6-s EEGs and the different subbands (delta, theta, alpha, beta, and gamma) of EEGs for the identification of seizure and epilepsy through integration of wavelets [35], a signal processing technique, and chaos theory from nonlinear science [36]. Chaos theory is employed as

Conclusion

A novelty of this proposed model is being the first application of deep neural network for EEG-based seizure detection. A 13-layer deep learning CNN algorithm is implemented for the automated EEG analysis. An average accuracy of 88.7% is obtained with a specificity of 90% and a sensitivity of 95%. The performance (accuracy, sensitivity, and specificity) of proposed model is slightly lower than some of the works summarized in Table 4. The advantage of the model presented in this paper, however,

Conflict of interest

There is no conflict of interest in this work.

Acknowledgements

Authors thank Dr. Amir Adeli, Board-Certified Neurologist, Columbus, Ohio, for reviewing the manuscript several times from the perspective of neurologist and providing valuable clinical inputs which improved the paper considerably.

References (63)

  • World Health Organization

    Epilepsy

    (2017)
  • American Epilepsy Society

    Facts and Figures

  • Harvard Health Publications, Harvard Medical School

    Seizure Overview

    (2014)
  • International League Against Epilepsy (ILAE).dd...
  • D.Y. Ko et al.

    Epilepsy and Seizures

    (2016)
  • A. Krumholz et al.

    Quality standards subcommittee of the american academy of neurology

    Neurology

    (2007)
  • Y. LeCun et al.

    Deep learning

    Nature

    (2015)
  • A. Krizhevsky et al.

    ImageNet classification with deep convolutional neural networks

  • A. Ortiz-Garcia et al.

    Ensembles of deep learning architectures for the early diagnosis of alzheimer's disease

    Int. J. Neural Syst.

    (2016)
  • M.H. Rafiei et al.

    A novel machine learning model for estimation of sale prices of real estate units

    Construct. Eng. Manag.

    (2016)
  • M.H. Rafiei et al.

    Supervised deep restricted boltzmann machine for estimation of concrete compressive strength

    ACI Mater. J.

    (2017)
  • X. Glorot et al.

    Deep sparse rectifier neural networks

    J. Mach. Learn. Res.

    (2010)
  • I. Goodfellow et al.

    Deep Learning

    (2016)
  • J.G. Lee et al.

    Deep learning in medical imaging: general overview

    Korean J. Radiol.

    (2017)
  • R.G. Andrzejak et al.

    Indications of nonlinear deterministic and finite dimensional structures in time series of brain electrical activity: dependence on recording region and brain state

    Phys. Rev. E

    (2001)
  • N. Siddique et al.

    Computational Intelligence - Synergies of Fuzzy Logic, Neural Networks and Evolutionary Computing

    (2013)
  • K. Fukushima

    Neocognitron: a self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position

    Biol. Cybern.

    (1980)
  • M. Kallenberg et al.

    Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring

    IEEE Trans. Med. Imag.

    (2016)
  • S. Pereira et al.

    Brain tumor segmentation using convolutional neural networks in MRI images

    IEEE Trans. Med. Imag.

    (2016)
  • N. Hatipoglu et al.

    Cell segmentation in histopathological images with deep learning algorithms by utilizing spatial relationships

    Med. Biol. Eng. Comput.

    (2017)
  • J.H. Tan et al.

    Segmentation of optic disc, fovea, and retinal vasculature using a single convolutional neural network

    J. Comput. Sci.

    (2017)
  • Cited by (1188)

    View all citing articles on Scopus
    View full text