EEG analysis of seizure patterns using visibility graphs for detection of generalized seizures

Highlights

• A single-channel EEG signal is mapped into visibility graphs (VGS) so that network theory can be used.

• The power-law degree distributions in difference VG (DVG) show the best separation among the seizure EEG and non-seizure EEG.

• The connecting structure of horizontal VG (HVG) outperforms those of VG and DVG in distinguishing seizure EEG from non-seizure EEG.

• The proposed VGS-based features can help improve seizure detection for patients with intellectual disability.

Abstract

Background

The traditional EEG features in the time and frequency domain show limited seizure detection performance in the epileptic population with intellectual disability (ID). In addition, the influence of EEG seizure patterns on detection performance was less studied.

New method

A single-channel EEG signal can be mapped into visibility graphs (VGS), including basic visibility graph (VG), horizontal VG (HVG), and difference VG (DVG). These graphs were used to characterize different EEG seizure patterns. To demonstrate its effectiveness in identifying EEG seizure patterns and detecting generalized seizures, EEG recordings of 615 h on one EEG channel from 29 epileptic patients with ID were analyzed.

Results

A novel feature set with discriminative power for seizure detection was obtained by using the VGS method. The degree distributions (DDs) of DVG can clearly distinguish EEG of each seizure pattern. The degree entropy and power-law degree power in DVG were proposed here for the first time, and they show significant difference between seizure and non-seizure EEG. The connecting structure measured by HVG can better distinguish seizure EEG from background than those by VG and DVG.

Comparison with existing method

A traditional EEG feature set based on frequency analysis was used here as a benchmark feature set. With a support vector machine (SVM) classifier, the seizure detection performance of the benchmark feature set (sensitivity of 24%, FD_t/h of 1.8s) can be improved by combining our proposed VGS features extracted from one EEG channel (sensitivity of 38%, FD_t/h of 1.4s).

Conclusions

The proposed VGS-based features can help improve seizure detection for ID patients.

Introduction

Electroencephalography (EEG) is considered an important biometric for the diagnosis and screening of epileptic seizure detection in clinical practice. Many EEG-based seizure detection methods have been reported, aiming at differentiating non-seizure and seizure states for epileptic patients by using certain features extracted from EEG (Acharya et al., 2013, Alexandros et al., 2012, Ramgopal et al., 2014). These EEG features can be analyzed in different domains. For example, in the time domain, EEG amplitude, shape, and duration (Gotman et al., 1979), chaoticity or predictability measured by Lyapunov exponents (Stam, 2005), complexity evaluated by Lempel-Ziv complexity, as well as regularity by means of entropy measures (Acharya et al., 2012) have been studied. In the frequency domain, previous works have been addressing on characterizing epileptic seizures by means of EEG spectrum analysis (Subha et al., 2010, Polat and Gne, 2007). In addition, some methods have been applied to analyze EEG for seizure detection in the time-frequency domain (e.g. discrete wavelet transform (Saab and Gotman, 2005) and empirical mode decomposition (Alotaiby et al., 2014)) as well as in the spatial-temporal domain (e.g. phase locking synchrony (Mormann et al., 2000)).

In this work, we propose to analyze EEG signals in the ‘visibility’ domain, in which visibility algorithms can be used to characterize a time series signal. The concept of using the visibility algorithms to analyze time series was proposed by Lacasa et al. (2008). The visibility algorithm can be interpreted as a geometric transform of a periodic time series, which can be an analogy of the discrete Fourier transform (DFT) in the frequency domain. The visibility algorithms can capture the presence of nonlinear correlations of a time series such as chaotic behavior where DFT fails to capture (Luque et al., 2009). Among the visibility algorithms, a visibility graph (VG) method has been previously used to characterize one-dimensional signal by mapping it into a two-dimensional network in a graph a so that signal properties are geometrically visible (Lacasa et al., 2008). The VG methods have been applied to analyze electrophysiological signals such as ECG (Long et al., 2014) and EEG signals (Zhu et al., 2014), to exploit human sleep stages. For analyzing epileptic seizures, VG-based methods have been used for seizure detection based on the high-frequency sub-band of electrocorticography (ECoG) signals (Tang et al., 2013). The ictal and interictal EEG may be identified with the power law characteristics based on a VG method (Reijneveld et al., 2007). We thus evaluated the VG method in this study. In addition, we investigated the horizontal VG (HVG) (Luque et al., 2009), and difference VG (DVG) (Zhu et al., 2014), which are the extended versions of the basic VG. They are expected to capture different properties of EEG signals, e.g., HVG can better represent chaotic characteristics of EEG than VG (Luque et al., 2009), and DVG can help remove redundant information in VG (Zhu et al., 2014). The VG, HVG and DVG are termed as VGS.

The seizure detection (Wang et al., 2015, Wang et al., 2016, Wang et al., 2017) for a specific population with both epilepsy and intellectual disability (ID) is challenging based on existing EEG features due to the presence of abnormal EEG activities caused by cerebral development disorders (Steffenburg et al., 1998, Guerrini et al., 2001). Clinicians often encountered different types of EEG in ID patients, such as abnormal background EEG (slow activity, no alpha), frequent occurrence of focal anomalies, high levels of inter-ictal epileptic transients that resemble seizures, abnormal sleep and wake cycles (difficult to interpret sleep/drowsiness EEG), as well as different seizure discharge patterns from non-ID epileptic patients. These EEG abnormalities may affect the seizure detection performance. It is because that seizure detection is to detect the abnormal ictal EEG, while the increased amount of abnormal interictal EEG causes more interference to the task of seizure detection. For example, an association has been found between intractability and abnormal EEG background in childhood epilepsy (Ko and Holmes, 1999). The VGS-based features can provide supplementary ‘visibility’ information to the existing EEG features in the conventional time and frequency domain. Therefore, they might potentially benefit the seizure detection in ID patients. In addition, EEG signals vary across individuals, which often limits the application of a seizure detector in a large population. The variance of EEG signals in ID patients is often larger than that in non-ID patients, due to different levels of brain development. Our current study suggests that the ID population is a heterogeneous entity, causing significant variance of detection performance across subjects. However, when a one-dimensional EEG signal is mapped into a two-dimensional VGS, the geometrical properties of VGS is more robust to the variance in the original time and frequency space. For example, a mapped VG network can remain less variant when the original EEG signal has higher variances such as horizontal rescaling (i.e., changed frequency), vertical rescaling (changed amplitude) and addition of a linear trend (baseline drift) (Lacasa et al., 2008). This property of VGS-based methods is expected to reduce the variability of extracted EEG features within and across subjects. We thus evaluate the feasibility of the VGS-based features on seizure detection in the ID population.

Most of previous automatic seizure detection studies consider seizure detection as a classification task between the seizure class and non-seizure class. However, within the seizure class, the ictal EEG generally shows various morphological patterns. We term such ictal morphological patterns as the EEG seizure patterns (or seizure patterns), also known as epileptiform discharges (Ko, 2016) or polymorphic seizure patterns (Meier et al., 2008). Note that EEG seizure patterns differ from the clinical seizure types defined by ILAE (Fisher et al., 2017), which are diagnosed based on both EEG and clinical symptoms. The seizure patterns have been shown to associate with seizure detection performance (Wang et al., 2017, Meier et al., 2008). Typical seizure patterns include spikes, spike-wave complexes, and sharp/slow waves (De Lucia et al., 2008). In addition, the seizures accompanied by electromyography (EMG) artifacts are common in this ID population, and it is not appropriate to simply exclude them as artifacts (Conradsen et al., 2011). Therefore, we define them as the EMG seizures (Wang et al., 2017). Fig. 1 illustrates the four typical seizure patterns in the ID patients. These patterns associate with clinical seizure types defined by ILAE. Specifically, fast spikes exist in most tonic seizures; spike-wave patterns often occur during absence-like seizures or at the end of tonic–clonic seizures; slow waves may present during focal seizures and rhythmic delta/theta seizures; and seizure-related EMG can exist in most tonic, tonic–clonic and myoclonic seizures. The spike-waves and rhythmic waves can also occur in interictal EEG with shorter durations (Wang et al., 2017). The four seizure patterns may occur in a sequence or as a combination (e.g., polyspike complexes) during a seizure or non-seizure EEG. The EEG analysis in the visibility domain may enlighten its applicability in the recognition of seizure patterns, thereby facilitating the diagnosis of clinical seizure types.