Electrical status epilepticus during sleep electroencephalogram waveform identification and analysis based on a graph convolutional neural network

https://doi.org/10.1016/j.bspc.2022.103788Get rights and content

Abstract

Electrical status epilepticus during sleep (ESES) is an epileptic syndrome in which neurons in the brain continue to discharge during the sleep phase and is common in mid-childhood. Affected patients often experience a decline in cognition, learning ability, memory, and expressive language skills. Therefore, timely and accurate diagnosis can effectively protect the health of a patient. Currently, the identification and analysis of ESES activities mainly rely on manual detection or traditional matching learning algorithms, such as morphology and template matching. These algorithms are time-consuming or have low accuracy. Therefore, in this paper, we propose a graph convolutional neural network that can automatically and accurately identify ESES activity from non-ESES activity. We divide the whole EEG signal into small segments, each of which covers one second of the EEG data. Then, we construct a graph according to each segment of the EEG data and train a graph convolutional neural network to classify the graph into two categories: ESES or non-ESES. Compared with other state-of-the-art algorithms, for the proposed algorithm, the accuracy, F1-Score, Area Under Curve(AUC) and sensitivity reaches 91.2%, 95.0%, 96.5%, and 91.3%, respectively, and outperforms the other algorithms.

Introduction

According to the statistics from the World Health Organization (WHO), about 50 million people worldwide suffer from epilepsy, with about 2.4 million people diagnosed each year. The risk of premature death in epilepsy is three times higher than that of the general population [1]. ESES is a type of epilepsy symptom which is common in children aged 5–10 years, and its inducement is very complex. Compared with conventional epilepsy patients, ESES patients often have severe developmental problems [2], [3]. Patients often have seizures during sleep, some accompanied by muscle clonus or partial convulsions, and some without any specific manifestations [4] which also bring many obstacles to doctors’ diagnoses. This leads to delays in treatment and management, delaying the best time for treatment. It is, therefore, crucial to reduce the risk of future seizures and epilepsy-related complications as well as to keep patients safe from illness. It has also been shown [5] that two-thirds of patients with epilepsy can be controlled with antiepileptic drugs. Doctors usually need to assess the frequency of patients’ seizures and use different drug doses accordingly. There are many medical instruments capable of measuring EEG, in contrast, EEG is not only inexpensive but also accurately responds to the activity of different brain regions and thus is often used for a diagnosis. For ESES patients, the main manifestations in an EEG are continuous abnormal spikes and slow waves [6]. Clinically it is often necessary to determine the seizure frequency of patients within a certain period based on an EEG and to use different doses of drugs according to this frequency. However, the current clinical recognition of epileptic signals mainly relies on manual recognition by physicians, requiring them to have rich experience in EEG recognition. But for long-term EEG, the subject's EEG testing time is often up to tens of hours, and the use of manual testing will greatly increase the burden on the physician. Therefore, in practice, clinicians often do not read through all the EEG records, but only consider a few minutes of them for analysis. The final seizure frequency of the overall data is predicted based on this small portion of data, thus some errors may occur. Second, there may be differences between different experts, which reduces the diagnostic efficiency. Finally, for manual recognition, the requirements on the experience and professionalism of the physician are extremely high. Given the abovementioned shortcomings of manual detection, the need for ESES automatic detection algorithms in the clinical setting is necessary and urgent. The main current algorithms for the detection of ESES activity are template matching [7], [8], [9], [10], morphology [11], [4], wavelet transform [12], traditional machine learning [13], artificial neural networks [14], [15], and deep learning [15], [16].

In [7], a template matching algorithm was proposed to detect typical spikes in ESES activity by calculating the number of interrelationships between EEG and template spikes. If this coefficient exceeds a certain threshold then it is classified as a spike. The work in [8] combines template matching with a clustering algorithm to achieve the classification of ESES activities and non-ESES activities. However, the template matching algorithm requires a predetermined template, and the design of the template often leads to a decrease in sensitivity by ignoring some atypical spike features. The variability and diversity of waveform features among patients also make the template matching method difficult to apply in practical situations.

In [4], the authors quantified epileptiform activity in ESES based on the identification of spike and slow-wave abnormalities and the estimation of their durations by first setting morphological filters to shape the EEG signals to obtain candidate spikes and slow waves. Following this, spikes and slow waves are extracted from the candidate spikes and slow waves by setting thresholds and predicting their onset and offsets. However, this method does not take into account the diversity of patients, and secondly, the parameters in this method, e.g., morphological filters, thresholds, etc., rely heavily on specialization, resulting in a model with limited generality.

All of the above algorithms have difficulty in achieving the automatic detection of epileptic activity, and their final results are often related to the parameters set by human experience. The work in [13] combines the a dual-threshold wavelet transform with machine learning algorithms to obtain candidate spikes, extracts the amplitude, slope, and peak during this period, and classifies these features via an SVM to determine whether they are ESES activities. Although the above automatic detection method based on machine learning achieves high accuracy, its drawbacks are also very obvious. First, machine learning has some advantages in learning from small data samples but it is weak for large data samples. Second, machine learning models are simple, while seizures are highly nonlinear in EEG and dynamically change in time series. There are large differences between patients, therefore traditional machine learning models have difficulty in learning hidden features. These shortcomings also lead to the weak generalization ability of the model, which makes it difficult to be applied clinically.

With the constant iterations of technology, neural networks have emerged. The work in [16] first performed a wavelet transform on the original EEG signal, then extracted 280 features from the four sub-bands of the signal in both time and frequency domains, and finally sent these features as a dataset into a deep belief neural network for training to identify ESES activity. The work in [14] used a fuzzy artificial neural network algorithm to classify ESES activity and non-ESES activity. First, the dataset was divided into experimental and control groups, the experimental group comprised of ESES patients, while the control group comprised of normal people. Then, 100 segments of ESES activity and 100 segments of non-ESES activity were selected from these two data groups, each segment was 8 s in length, and for each segment, the ranking entropy and sample entropy were fed into the fuzzy artificial neural network as features for classification. The final classification accuracy reached 89.0%. However, this method was only validated on a dataset with a total duration of 1 h, which did not prove the generalization of this method. Also, its experimental results contained the training set and were not validated on the test set alone. Finally, the fuzzy neural algorithm was able to perform well on small sample data but has limited ability to learn for large numbers of complex samples. In [15], the authors selected 1500 spikes with 150,000 background waves in each patient's EEG signal, each with a length of 0.5 s. The selected signals were then fed into a convolutional neural network consisting of a convolutional layer with a fully connected layer to achieve the recognition of epilepsy and non-epilepsy. First, in this study, the selection of spikes and background waves introduces a lot of human factors, and the selected ones are often typical epileptic activities. However, seizures are often very complex and the waveforms of seizure periods are diverse, which may lead to a limited recognition performance of the model. Second, the constructed model is relatively simple, which will have difficulty in coping with the recognition of waveforms in actual complex situations.

In recent years, researchers have found that there are specific relationships between signals from different regions in the brain [17]. Brain disorders are rarely confined to a specific region,instead, there is often a propagation along axonal pathways in the brain [18]. Brain networks are often represented by graph structures, which are non-Euclidean structures. Traditional CNNs are generally used to deal with planar-type Euclidean structures, and to fully exploit the spatial relationships in brain networks, graph convolutional neural networks were proposed [19]. Graph convolutional neural network models are widely used for emotion recognition [20], [21] sleep stage classification [22], and motor image recognition [23]. However, due to the diverse seizure characteristics and complex seizure mechanisms of epileptic EEG signals, there are fewer studies using graph convolutional neural network models in epilepsy compared to their applications in other fields. In this paper, a Chebyshev polynomial-based graphical convolutional neural network is used for ESES detection, and the main contributions of this paper are as follows:

  • 1.

    Since epilepsy data are non-Euclidean structures from the brain network perspective, it is not possible to perform convolution operations using conventional operators, as these do not satisfy shift-invariance and shared connectivity. Therefore, this paper proposes to use a Chebyshev polynomial-based graph convolution neural network and extracts features for each channel of EEG signals from both frequency and time domains separately. These features are then used as node features of the graph data.

  • 2.

    Traditional convolutional neural networks convert EEG data into two- or three-dimensional grid data, destroying the original data structure. Our proposed graph convolutional neural network relies on the a priori knowledge of the relationship between the graph and the spatial structure of the brain, preserving the original graph data structure and thus providing a strong learning capability for ESES detection.

  • 3.

    Compared with the previous work [14] with 8 s as the classification granularity, this paper can achieve a fine-grained classification of 1 s, which is more difficult and more demanding for the model.

  • 4.

    While most previous work on the identification of ESES activity was based on a small sample of datasets, the proposed model in this paper demonstrates its generalization ability by validating the identification of ESES across a large number of different patients.

Section snippets

Datasets

The dataset used in our study was provided by the Children’s Hospital of Chongqing Medical University. All patients were infants and children during sleep, and electrode sheets were placed on the patient's scalp in the international standard 10–20 placement method. The dataset for our project contains 49 patients, each subject has an EEG duration of 1 h, for a total of 49 h, of which the training set has a total of 122,366 s, the seizures in the training set are a total of 114,082 s, and the

Conclusions

According to the above analysis, the graph convolutional neural network used in our work not only outperforms the current common models in terms of evaluation metrics, but also has a smaller size and consumes fewer computer resources. The experimentation demonstrates that our proposed model can achieve better results for different patients and can use fewer graph convolution parameters with the a priori nature of the graph. Therefore, the graph convolutional neural network used in our work has

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Lu Meng reports was provided by Northeastern University. Lu Meng reports a relationship with Northeastern University that includes: board membership. Lu Meng has patent pending to None. None

Acknowledgment

This work was funded by the National Key R&D Program of China under Grant 2018YFB2003502, National Natural Science Foundation of China (62073061), and the Fundamental Research Funds for the Central Universities (N2204009).

References (32)

  • R. San-Segundo et al.

    Classification of epileptic EEG recordings using signal transforms and convolutional neural networks

    Comput Biol Med

    (2019)
  • D.K. Hammond et al.

    Wavelets on graphs via spectral graph theory

    Appl. Comput. Harmon. Anal.

    (2011)
  • W.H. Organization

    Epilepsy [Online]

    (2019)
  • Samanta D, A.K.Y. (2021). Electrical Status Epilepticus In Sleep [Online]. Available:...
  • Zhao, X., Wang, X., Chen, C., Fan, J., Yu, X., Wang, Z., et al. (2020). A knowledge-based approach for automatic...
  • M.M. Goldenberg

    Overview of drugs used for epilepsy and seizures: etiology, diagnosis, and treatment

    P & T : a peer-reviewed journal for formulary management

    (2010)
  • Cited by (0)

    View full text