Selection of features for patient-independent detection of seizure events using scalp EEG signals
Graphical abstract
Introduction
Chronic lesion epilepsy induces transient brain dysfunction via the abnormal discharge of brain neurons [1,2]. Epilepsy is recognized as the second most common neurological disorder, ranked just after stroke [3]. More than 20% of epilepsy patients suffer from seizures, which are usually refractory to medication [4]. Epileptic seizures occur recurrently and unpredictably, and substantially impact the quality of life of the patient [5,6]. Various modern imaging technologies have been applied to diagnose epilepsy, including magnetic resonance imaging (MRI) [7,8], computed tomography (CT) [9], positron emission tomography (PET) [10,11] and electroencephalogram (EEG) [12].
EEG technology records electrophysiological signals to describe brain activity, and has been widely used to detect epilepsy [5,12]. Although EEG has characteristics of low cost, non-invasiveness, and high signal resolution [13,14], clinical annotation of EEG signals still relies heavily on human screening [15]. The visual inspection of EEG signals is recommended by guidelines, but is tedious and time-consuming [16].
Machine learning algorithms have become the research focus of EEG-based epileptic seizure detection problems [[17], [18], [19]], and several studies have investigated classification based on EEG signals [20,21]. Epileptic seizure detection is formulated as a classification problem of ictal and interictal EEG signals; various feature extraction algorithms have been utilized to generate features from a fixed-length window of EEG signals, including the time domain [22], frequency domain [23], and non-linear characteristics [24,25]. After the feature extraction step, a variety of feature selection [26,27] and classification algorithms [5,28] may be utilized to generate seizure detection models. Deep learning algorithms such as CapsNet and a deep belief network may also be utilized to characterize the phenotypic patterns within EEG signals [29,30].
Most studies calculate variables from EEG signals using feature extraction algorithms, and formulate EEG-based seizure detection as a machine learning problem using these extracted features [19,31,32]. Gabor et al. trained a self-organizing map (SOM) neural network to detect seizures in 24 long-term EEG recordings using time-frequency features, and achieved an accuracy of 90%, thus demonstrating the possibility of automated seizure detection [33]. Time-frequency features are widely used to describe EEG signals due to their non-stationary characteristics. Various forms of wavelet transformations have been utilized to extract time-frequency features from EEG signals. Saab and Gotman used wavelet decomposition to extract features from scalp EEG signals and achieved a sensitivity of 76.0% for the detection of seizures [34]. Discrete wavelet transformation (DWT) is a powerful tool for analyzing time-frequency domain data, and can give a more flexible representation of signals. An extreme learning machine (ELM) was used to train a seizure detection model in Ref. [35]. In addition to time-frequency features, non-linear features can also be extracted from EEG signals. Kannathal et al. investigated the seizure detection problem by integrating four entropy estimators: Shannon spectral entropy, Renyi's entropy, Kolmogorov-Sinai entropy and approximate entropy. Their entropy-based model achieved a seizure classification accuracy of 90% [36]. Another study demonstrated that the non-linear features such as correlation dimension (CD), largest Lyapunov exponent (LLE), Hurst exponent (H), and entropy were accurate when used for seizure detection [37].
This study proposes a novel feature extraction algorithm in the time domain for the patient-independent seizure detection (pidSeizure) problem, and demonstrates that the integrated modeling of various feature sources can achieve satisfactory detection performance. Although much progress has been achieved in the epileptic seizure detection problem, most prior studies have focused on training a separate detection model for each patient. In this study, we hypothesize that a pidSeizure model might better serve epileptic seizure patients. Our experimental data also suggest that the extracted features can be further selected for better classification performance. The final detection model utilizes both feature extraction and feature selection to deliver accurate seizure detection performance.
Section snippets
Problem setting
This study investigated the patient-independent detection problem of epileptic seizures using scalp EEG signals, due to the following two challenges identified in existing studies.
Previous studies have usually used patient-specific modeling [21,24,[38], [39], [40]], i.e., they have trained a separate model for each patient. Although this approach can achieve high detection accuracy, clinicians must wait for several epileptic seizure events before a new patient can be monitored. In addition, the
Extraction of features from raw EEG data
Equal numbers of positive and negative samples were collected from each patient for further analysis, as shown in Table 1. Hence, the number of sample pairs collected from each patient was the minimum of the numbers of positive and negative samples for this patient. A total of 1260 pairs of samples were collected from all patients.
Two families of features were extracted from each channel of a given EEG sample, i.e., the time domain and the nonlinear domain, as shown in Table 2. We calculated 21
Conclusions
Epileptic seizure is a nervous system disorder that does not directly cause lethal damage to patients; however, seizures can substantially impact the quality of life of patients, and can give rise to major mental pressures. The accurate detection of seizures may provide vital information to help both patients and their family members to maintain their daily routines and quality of life.
This study focused on the patient-independent seizure detection problem, which is more challenging to solve
Data availability
The dataset used in this experiment was downloaded on May 4, 2018, and can be obtained from https://www.physionet.org/content/chbmit/1.0.0/.
Declaration of competing interest
None Declared.
Acknowledgements
This work was supported by the Jilin Provincial Key Laboratory of Big Data Intelligent Computing (20180622002JC), the Education Department of Jilin Province (JJKH20180145KJ), and the startup grant of the Jilin University. This work was also partially supported by the Bioknow MedAI Institute (BMCPP-2018-001), the High Performance Computing Center of Jilin University, and by the Fundamental Research Funds for the Central Universities, JLU. The insightful comments from the three anonymous
References (108)
- et al.
Targeting of microRNA-21-5p protects against seizure damage in a kainic acid-induced status epilepticus model via PTEN-mTOR
Epilepsy Res.
(2018) - et al.
Patient-specific seizure onset detection
Epilepsy Behav. : E&B
(2004) - et al.
Epilepsy diagnosis: positron emission tomography
Handb. Clin. Neurol.
(2012) - et al.
Dynamic imaging of ictal oscillations using non-invasive high-resolution EEG
Neuroimage
(2011) - et al.
Fusion of facial expressions and EEG for implicit affective tagging
Image Vis Comput.
(2013) - et al.
Autoencoding of long-term scalp electroencephalogram to detect epileptic seizure for diagnosis support system
Comput. Biol. Med.
(2019) - et al.
Classification of epileptic EEG recordings using signal transforms and convolutional neural networks
Comput. Biol. Med.
(2019) - et al.
Automated epileptic seizure detection based on break of excitation/inhibition balance
Comput. Biol. Med.
(2019) - et al.
Sleep spindle detection using artificial neural networks trained with filtered time-domain EEG: a feasibility study
Comput. Methods Progr. Biomed.
(2005) - et al.
Automated seizure detection using limited-channel EEG and non-linear dimension reduction
Comput. Biol. Med.
(2017)