EEG-Based Seizure detection using linear graph convolution network with focal loss

https://doi.org/10.1016/j.cmpb.2021.106277Get rights and content

Highlights

  • We propose to use LGCN to exploit the spatial relationship between EEG electrodes to enhance the representation power of EEG signals during normal brain activities and seizures.

  • The correlation of different electrodes are incorporated in the feature learning process, which improves the seizure detection performance.

  • The proposed approach yields superior performance over the-state-of-the-art in seizure detection tasks on the CHB-MIT dataset.

Abstract

Background and Objectives: Epilepsy is a clinical phenomenon caused by sudden abnormal and excessive discharge of brain neurons. It affects around 70 million people all over the world. Seizure detection from Electroencephalography (EEG) has achieved rapid development. However, existing methods often extract features from single channel EEG while ignoring the spatial relationship between different EEG channels. To fill this gap, a novel seizure detection model based on linear graph convolution network (LGCN) was proposed to enhance the feature embedding of raw EEG signals during seizure and non-seizure periods. Method: Pearson correlation matrix of raw EEG signals was calculated to build the input graph of the graph neural network where the coefficients of the matrix models the spatial relations in EEG signals. The last softmax layer makes the final decision (seizure vs. non-seizure). In addition, focal loss was utilized to redefine the loss function of LGCN to deal with the data imbalance problem during seizure detection. Results: Experiments are conducted on the CHB-MIT dataset. The seizure detection accuracy, specificity, sensitivity, F1 and Auc are 99.30%, 98.82%, 99.43%, 98.73% and 98.57% respectively. Conclusions: The proposed approach yields superior performance over the-state-of-the-art in seizure detection tasks on the CHB-MIT dataset. Our method works in an end-to-end manner and it does not need manually designed features. The ability to deal with imbalanced data is also attractive in real seizure detection scenarios where the duration of seizures is much shorter than the lasting time of non-seizure events.

Introduction

Epilepsy is a clinical phenomenon caused by sudden abnormal and excessive discharge of brain neurons. Epileptic seizures are usually systemic seizures, which can make people loss consciousness and seriously affect people’s normal life. Epilepsy diagnosis can be supported by EEG signals which are collected through electrodes placed on the scalp or cerebral cortex and carries pathological information related to the brain [1], [2]. When EEG signals are expressed as background wave abnormalities or epileptic discharge, it can be diagnosed as epilepsy [3], [4].

In recent years, seizure detection from EEGs has made rapid progress. Coan started the early seizure detection work [5]. Chaos theory provided new ideas for seizure detection and prediction [6]. The original EEG signal was decomposed into many basic signal waves and the slope and amplitude value were taken as the main characteristics of seizure detection. Dong et al. presented to use sparse representation to analyze the epileptic EEG signals and cognitive impairment [7]. Wang presented a large scale integration architecture of three-class classification for epilepsy and seizure detection [8]. In [9], time-frequency features are extracted to identify seizures in EEG signals. Das et al. used variational mode decomposition and mode spectral entropy to detect seizure events in EEG signals [10]. The extracted features were used to classify normal or abnormal EEG signals.

Deng et al. presented a transductive joint knowledge transfer method to recognize epileptic EEG signals [11]. In the work of Chen et al., an efficient, automatic and individualized approach was developed for preictal and interictal stage identification [12]. Discrete wavelet transform (DWT) was proposed for seizure detection. Gill et al. detected epileptic seizures from EEG signals using hybrid feature set [13]. By simultaneously capturing the main information, the model can learn invariant seizure representations. This approach was superior to previous approaches in terms of sensitivity and false positive rate on across-patient classifiers. Aggelos decomposed EEG recordings using linear discriminant analysis (LDA) and naive Bayesian. Seizures were detected in patient specific manner through Poincare section [14].

Besides the above mentioned methods, deep learning has gained rapid attention in EEG signal processing. Peng et al. proposed dictionary learning and sparse representation to detect neonatal seizures [15]. Hossain used convolutional neural network (CNN) to automatically select optimal feature subsets in EEG [16]. Multiple features were taken as the input of the introduced CNN, and the optimal features were the corresponding output. Similarly, Ansari et al. used a deep CNN with random forest (RF) to automatically select optimal feature subsets in seizure detection in EEG [17]. More recently, different architectures of long short-term memory (LSTM) [18], [19] and CNN [20], [21] have been proposed to detect and predict seizure events from EEG signals.

Most of the methods mentioned above rely on the single channel EEG while ignoring the spatial relationship of different electrodes, which may lead to the loss of important information. Ian et al. took the structural relationship into consideration and proposed a temporal graph convolutional network (TGCN) model to detection seizures from EEG signal [22]. Zeng et al. utilized a hierarchy graph convolutional network (HGCN) model for seizure detection and classification [23]. However, the works do not tackle the class-imbalance problem in seizure detection. Sample imbalance is common in seizure detection where the duration of seizures is much shorter than the lasting time of non-seizure events. Few works have been proposed to tackle the second problem in EEG based seizure detection. Yuan et al. detected seizures using wavelet packet transform and dealt the imbalanced problem with extreme leaning machine (ELM). Yuan et al. proposed an imbalanced seizure detection framework based on EasyEnsemble learning [24]. But the feature extraction and classification processes were separated. The manually designed features may not be optimal.

In order to cope with the above mentioned two problems, a novel seizure detection model based on linear graph convolution network (LGCN) [25] was proposed to enhance the representation of raw EEG signals during seizure and non-seizure periods. The whole model works in an end-to-end manner. Pearson correlation matrix of raw EEG signals was calculated to build the graph. The last softmax layer makes the final decision (seizure vs. non-seizure). Further more, focal loss [26] was utilized to redefine the loss function of LGCN to deal with the data imbalance problem during seizure detection. The proposed model achieves superior seizure detection performance on the CHB-MIT dataset. The main contributions of the paper are as listed follows:

  • 1.

    We propose to use LGCN to exploit the spatial relationship between EEG electrodes to enhance the representation power of EEG signals during normal brain activities and seizures.

  • 2.

    The correlation of different electrodes are incorporated in the feature learning process, which improves the seizure detection performance.

  • 3.

    The proposed approach yields superior performance over the-state-of-the-art in seizure detection tasks on the CHB-MIT dataset.

The rest of the paper is organized as follows. Section 2 illustrates the network architecture, the learning process and the focal loss function of the proposed model. Section 3 elaborates experiments, including the evaluation dataset and indicators, ablation experiments and comparisons with state-of-the-art methods. Section 4 gives the discussions and Section 5 concludes the paper.

Section snippets

General introduction

Fig. 1 shows the flow chart of the proposed LGCN model for seizure detection. The overall framework of the model includes three parts. The first part is data acquisition and pre-processing. The second part is correlation analysis and graph construction. Finally, the raw EEG data together with the graph are fed into the LGCN networks. After fully connection operation, the last softmax layer gives the final decision (seizure or non-seizure).

Data pre-processing

EEG signals collected from the international 10–20

Experimental dataset

Experiments are conducted on CHB-MIT dataset [37]. The dataset is collected from patients in Boston Children’s Hospital. There are 5 males (aging from 3 to 22) and 17 females (aging from 1.5 to 19). The signals are sampled with a frequency of 256 Hz. Each EEG recording lasts about one hour. The total duration of the signals is 958 hours long, of which 198 hours are seizure events.

The International 10–20 system was exploited to collect these signals. Our experiments use the common EEG signals

Discussion

The CHB-MIT dataset has been extensively evaluated in many works [10], [11], [12], [13], [14], [15], [16], [38], [39], [40], [41], [42], [43]. The comparison of the proposed LGCN+focal loss model and other methods are given in Table 4. From Table 4 we can see that our method outperforms most of the compared seizure detection methods. In particular, our method is 6.54% higher in specificity than wavelet transform [12] and 10.55% in sensitivity higher than LDA [14]. This is because most of the

Conclusion

In this paper, we propose a LGCN+focal loss model to detect seizures from raw EEG data. The LGCN part of the model takes the channel relationship into consideration and treats each EEG channel as a node in the graph. The topological relationships of the channels are modeled by the correlation coefficients among them. In order to deal with the problem of data imbalance in EEG based seizure detection, the loss function of LGCN is re-defined using the focal loss function. The new loss function

Declaration of Competing Interest

The authors declare that they have no conflict of interest.

Acknowlgedgments

This work was made possible through support from the China Postdoctoral Foundation (No.2017M612335), National Natural Science Foundation of China (NO.81871508, No.61773246, No.61701270, NO.61501283), the program for Youth Innovative Research Team in University of Shandong Province (NO. 2019KJN010).

References (43)

  • D. Wen et al.

    Review of sparse representation-based classification methods on eeg signal processing for epilepsy detection, brain-computer interface and cognitive impairment

    Frontiers in Aging Neuroence

    (2016)
  • Y. Wang et al.

    Hardware design of multiclass svm classification for epilepsy and epileptic seizure detection

    IET Circuits, Devices & Systems.

    (2018)
  • G. Carlos et al.

    New feature extraction approach for epileptic eeg signal detection using time-frequency distributions

    Medical & Biological Engineering & Computing

    (2010)
  • D. Priya et al.

    Detection of epileptic seizure event in eeg signals using variational mode decomposition and mode spectral entropy

    2018 IEEE 13th International Conference on Industrial and Information Systems (ICIIS)

    (2018)
  • D. Chen et al.

    A high-performance seizure detection algorithm based on discrete wavelet transform (dwt) and eeg

    Plos One.

    (2017)
  • A. Gill et al.

    Analysis of eeg signals for detection of epileptic seizure using hybrid feature set

    Theory and Applications of Applied Electromagnetics

    (2015)
  • K. Aggelos

    Analysis of high-dimensional phase space via poincare section for patient-specific seizure detection

    IEEE transactions on neural systems and rehabilitation engineering: a publication of the IEEE Engineering in Medicine and Biology Society

    (2016)
  • P. Hong et al.

    A novel automatic classification detection for epileptic seizure based on dictionary learning and sparse representation

    Neurocomputing

    (2019)
  • M. Hossain et al.

    Applying deep learning for epilepsy seizure detection and brain mapping visualization

    ACM Trans. Multimedia Comput. Commun. Appl.

    (2019)
  • A. Amir et al.

    Neonatal seizure detection using deep convolutional neural networks

    Int J Neural Syst

    (2018)
  • Y. Li et al.

    Automatic seizure detection using fully convolutional nested lstm

    Int J Neural Syst

    (2020)
  • Cited by (44)

    View all citing articles on Scopus
    View full text