Epileptic seizure focus detection from interictal electroencephalogram: a survey

Epilepsy, one of the most common neurological disorders, can affect people of any age, race, or ethnic background. According to the latest study by the World Health Organization (WHO), approximately 65 million people worldwide are affected by epilepsy, and there are an estimated 2.4 million new cases each year (Giannakakis et al. 2014; Levesque et al. 2017; Stafstrom and Carmant 2015). Epilepsy, defined as repeated and unpredictable seizures, causes social impairment and a high risk of death (Fisher et al. 2014; Pati and Alexopoulos 2010). Childhood epilepsy also seriously impacts the development of the brain by reducing learning ability and mental growth. Epileptologists generally classify seizures as either focal or generalized based on abnormal brain activities (Ngugi et al. 2011; van Mierlo et al. 2014). To control epileptic seizures, epileptologists prescribe anti-epileptic drugs. When these medicines fail to control the seizures, surgical removal of the epileptic focus may be the patient’s best chance for seizure freedom.

According to the clinical guidelines related to epilepsy surgery, the epileptic seizure focus is the cortex area from which the seizures originate (Lüders et al. 2006). The identification and surgical removal of the focus must be resected (inactivated or completely disconnected) for complete seizure freedom. Standard diagnostic methods include inspection of seizure semiology, high-resolution magnetic resonance imaging (MRI), and EEG. The scalp EEG is a non-invasive method of recording electrical activity by placing electrodes on the scalp using the international standard 10–20 system (Paul 2018). It provides one of the promising ways to identify the epileptic seizure focus before surgical intervention. However, when epileptologists cannot determine an epileptic seizure focus using non-invasive methods, they indicate to use intracranial EEGs (iEEGs) with the implantation of intracranial electrodes during both interictal and ictal phases. Before the epileptic focus resection for complete abolition of seizures, the epilepsy surgeon should consider integrating the multi-channel intracranial electrodes to these “area of cortex” and recording iEEG signals until collecting enough data from habitual seizures to analyze. The epileptologists then need to inspect the multi-channel iEEG data to identify seizure onset zones (SOZs) from within the recorded cortex area. During these inspections, the epileptologists need to analyze and label all long-term multi-channel iEEG data, the manual detection of which is challenging and time-consuming. The success of the epileptic focus resection for seizure freedom depends on accurate detection of the seizure focus. A key to achieving good results with resection surgery is identifying and resecting the area that may cause seizures. Such areas are called epileptogenic zones (EGZs). Currently, there is no single, non-invasive test method that can identify those areas. Approximately 20–30 percent of patients suffer from recurrent seizures after surgery (Elsharkawy et al. 2011).

In the endeavor to design computer-aided diagnosis tools, both non-invasive and invasive iEEGs are promising procedures. Automatic detection of seizure focus is highly desired, as it would reduce the epileptologist’s workload and would, along with other tests, increase confidence in related medical decisions. Also, computer-aided and data-driven approaches may provide a way of revealing a mechanism of epileptogenesis.

For designing the computer-aided systems based on the different types of EEG modalities, Some recent studies used biomarkers, including high-frequency oscillations (HFOs)(Zijlmans et al. 2011; Jacobs et al. 2009; Urrestarazu et al. 2007), phase-amplitude coupling (PAC)(Guirgis et al. 2015; Motoi et al. 2018; Amiri et al. 2019), interictal epileptiform discharges (IEDs) (Staley and Dudek 2006; Elsharkawy et al. 2011) while others utilized feature-extraction methods (Sharma et al. 2015b; Akter et al. 2020a, 2019; Itakura and Tanaka 2017). In biomarker-related studies to identify epileptic seizure focus, the computer-aided solutions have combined the epilepsy biomarkers in EEG signals with advanced signal and machine-learning approaches. The epilepsy biomarkers in EEGs are essential for identifying the epileptic seizure focus within conventional clinical systems. However, from a machine learning perspective, the features extracted from EEG signals to characterize epileptic signals are also good candidates for identification of epileptic focus. To identify the epileptic focus from EEG signals, the recorded EEG signals are typically pre-processed using various signal processing methods, and features are extracted from the pre-process EEGs signals to represent in a compact form (Akter et al. 2020a, 2019; Hassan et al. 2019). The different studies introduced feature-extraction methods to design the computer-aided solutions for identifying epileptic focus (Sharma et al. 2015b; Akter et al. 2020a, 2019; Hassan et al. 2019; Itakura and Tanaka 2017). Based on the biomarker and feature-extraction related studies, there is a large diversity of research that have been explored in the last years to design AI-systems. Many new algorithms have been designed and explored in order to identify epileptic seizure focus.

However, there are few review papers that surveys works on automated machine learning-based focus detection methods. Recently, one review paper for focal and non-focal epilepsy localization (Hussein et al. 2018) has summarized the several studies with a variant of datasets to identify focal and non-focal channels. However, the main goal of most studies reported in their survey was to identify seizure patterns in continuous EEG signals instead of localizing focal and non-focal channels used in Bern Barcelona dataset (Andrzejak et al. 2001). For diverse research in the epilepsy, we would like to draw attention to the AI-solutions that different researchers have introduced for identifying the epileptic focus. In this paper, we survey the literature of the past decade to observe which new approaches have been investigated to design computer-aided automated systems and be the most efficient.

The epileptogenic zone (EGZ) is a conceptual definition of the area that causes seizures; resection of the EGZ halts them. When antiepileptic drugs fail to treat epilepsy, the patient is diagnosed with refractory epilepsy. One of the treatment approaches is removing the cortex’s possible area responsible for generating the abnormal neural activity resulting in epileptic seizure. In a study of surgical treatment of epilepsies, Lüders et al. (2006) categorized the EGZ as actual EGZ (commonly referred to SOZ) and potential EGZ. The study defined the SOZ as the cortex area from which actual clinical seizures are generated. The SOZ is most commonly localized by either scalp or intracranial EEG. In contrast, after resecting the SOZ, the seizure may continue from the other area of the cortex defined as the potential SOZ. This situation occurs when the surgery does not accomplish complete resection of the actual EGZ.

So far, the precise localization of the EGZ boundary cannot be characterized by any diagnostic methods. The most common methods to localize the SOZ are either scalp EEG or iEEG during both the interictal and ictal phases. Varied recent epilepsy research has suggested the use of ictal and interictal high-frequency activity (\(\ge 80\) Hz) to localize the epileptogenic zone. In clinical practice, the SOZ is used to localize the boundary of the EGZ (Rosenow and Lüders 2001). In defining the possible boundary of the EGZ, the different cortical zones are measured using some standard diagnosis methods. The descriptions of these cortical zones and diagnostic tests are provided in Table 1. During the presurgical evaluation, epileptologists should consider these cortical zones when estimating the location and extent of EGZ (Talairach and Bancaud 1966; Lüders et al. 2006). To localize the possible EGZ, Talairach and Bancaud (1966) introduced three zones, namely a lesional, an irritative, and an epileptogenic zone. They reported that the SOZ was a reliable index of the location and extent of the possible EGZ and defined the SOZ as the epileptogenic zone.

Decades later, Lüders et al. (2006) enriched that concept by defining five cortical zones: the irritative zone (generates interictal spikes), seizure-onset zone (initiates clinical seizures), the symptomatogenic zone, the lesion, and the functional deficit zone. The detailed presurgical evaluation always requires defining these zones to understand the spatial relationship with the epileptogenic zone. Lüders et al. (2006) observed that, for some patients, complete resection of the SOZ does not lead to seizure freedom. After resecting the SOZ, the epileptic seizures were generated from the area in the brain closest to SOZ. Their study further dichotomized the SOZ into the actual and potential SOZ and suggested that the complete resection of both the actual and potential SOZ may be necessary for complete seizure cessation (see in Fig. 1). Therefore, epilepsy surgeons and epileptologists need to define the location and the extension of the EGZ by observing both SOZs before epileptic focus resection. Since the location and extension of the EGZ is a theoretical concept based on defining the actual and potential SOZs, clinical experts may only decide about the EGZ if the patient after surgery is seizure-free. Bancaud and Talairach (1992) also observed that, with patients with medically-intractable epilepsy, the SOZ is involved in a single region and interconnected with distinct regions.

Research is still ongoing regarding identification of the SOZ and the extent of brain area that should be considered for resection in order to achieve successful surgical treatment of epilepsy. In conventional clinical procedure for identification of the SOZ ( both actual and potential), the EEGs with multiple channels are recorded until enough data on habitual seizures has been collected. An epileptologist visually inspects the long-term multi-channel EEG signals to identify the electrodes with the possible SOZ that must be resected in order to achieve freedom from seizures. However, this process is challenging and time-consuming for epileptologists. Several studies have recently proposed computer-aided solutions based on either epilepsy biomarkers (HFOs, PAC, and IEDs) or the use of feature-extraction methods in EEGs. In the rest of this paper, we have reviewed the computer-aided designs (CAD) that might be utilized to identify the possible SOZ (Table 2).

Public epilepsy datasets can be roughly divided into two categories: scalp EEG and intracranial EEG. The first is used to detect abnormalities related to epileptic seizure and the second to determine epileptic focus location. What follow are some of the most commonly-used scalp EEG datasets.Regarding the epileptic focus location problem, the most used public intracranial EEG dataset is the so-called Bern-Barcelona EEG dataset (Andrzejak et al. 2012). This dataset includes the iEEG signal recorded from five patients with long-standing pharmacoresistant temporal lobe epilepsy who are candidates for epilepsy surgery. iEEG signals are recorded by the AD-TECH device (Racine, WI, USA), which has a sampling rate of 512 Hz. The focal channel, which detected first ictal iEEG signal changes and judged throw visual by at least two clinical experts. The other channels are defined as non-focal channels. An example of focal and non-focal samples are shown in Fig. 2. The dataset includes 7500 samples of the focal and non-focal, respectively. Each sample is 20 seconds and is processed with the bandpass filter between 0.5 and 150 Hz by using fourth-order Butterworth filter.

This dataset has been extensively used as a benchmark in the literature; however, it also has the following drawbacks:

Apart from the above biomarker-related, computer-aided design (HFOs, PAC, and IEDs), the statistical features have, over the last couple of years, also been used as an engineering markers for the identification of possible SOZ. It is important to note that the iEEG is the biosignals, and such biosignals contain useful information of diseases about biological systems. However, it is not easy to identify that information solely through visual observation of the raw signals. EEG signals have a significantly low signal-to-noise ratio (SNR) and are frequently corrupted during recording by external- and internal-source artifacts (Islam 2015; Islam et al. 2018). The presence of artifacts and noise poses a significant challenge to analysis of recorded signals, and thus, to useful information extraction or classification. The direct use of machine-learning approaches to biosignals without pre-processing or feature extraction may lead to a decrease in sensible decisions by the automated system. Pre-processing, signal processing, and machine-learning approaches were combined with a standard computer-aided design to deal with EEG artifacts and to extract and decode the relevant features of EEG signals. In an epileptic focus detection system, the most commonly used pre-processing step was division of the long-term EEG data into a set of segments or frames. The segmentation of that long-term EEG data was dependent upon data size, system performance, and the direction of clinical experts (Akter et al. 2020a, 2019; Hassan et al. 2019).

Feature-extraction methods can be applied to the segments to extract the practical features. To implement epileptic focus detection, several studies since 2020 have suggested utilizing information-theoretic and statistical approaches. The statistical features have been shown to be particularly useful at detecting the epileptic focus in different studies. After extracting features from iEEG signals, the next step is to classify the normal and epileptic events using a classifier, such as k-nearest neighbor (kNN), linear discrimination analysis (LDA), support vector machine (SVM), or an implementation of a complex algorithms (i.e., neural networks (NN)).

For the design of computer-aided solutions to the problem of identifying epileptic seizure focus, several epilepsy research exploited Bern-Barcelona iEEG dataset, which was considered benchmark (Table 5). The Bern-Barcelona iEEG dataset consists of a pair of focal and non-focal signals recorded in the epileptogenic and non-epileptogenic zones of the brain. In order to observe the focal and non-focal channels in iEEG, Zhu et al. (2013) first introduced the delay permutation entropy (DPE) method with an SVM classifier. Experimental results show that the DPE index of focal channels was significantly lower than that of the non-epileptogenic hemisphere. Sharma et al. (2014) proposed sample entropy as a feature and the least square SVM (LS-SVM) classifier was used to discriminate focal and non-focal features. The mean and standard deviation (SD) were also used to differentiate between the two channels (Deivasigamani et al. 2016). Three machine-learning approaches, including kNN, fuzzy Sugeno classifier (FSC), and LS-SVM, have been trained using the wavelet entropy features (Sharma et al. 2015b). The features were Shannon, Renyi’s, Tsallis, fuzzy, permutation, and phase entropies. However, the nature of EEG and ECoG signals is very complex, non-stationary, and time-dependent. Due to the non-stationary properties of EEG data, the linear feature-extraction methods do not adapt perfectly to the EEG signals. Therefore, several studies proposed decomposition methods to improve the performance of the system, including wavelet transformation (WT) (Gupta and Pachori 2020), wavelet filter-bank (WFB) (Sharma et al. 2017), tunable-Q wavelet transformation (TQWT) (Sharma et al. 2017), flexible analytic wavelet transform (FAWT) (Gupta et al. 2017), and Empirical WT (EWT) (Bhattacharyya et al. 2018). A comparison study with Bern-Barcelona dataset using time-domain multiband analysis, including EMD and bivariate EMD (BEMD) (Rilling et al. 2007; Subasi et al. 2019), was reported by Itakura and Tanaka (2017) and (Bhattacharyya et al. 2018). Recently, new developments in computer-aided systems have suggested 23 feature-extraction methods that might be used to extract features from EEG signals (Acharya et al. 2019). The feature-extraction methods used to design those systems were different types of entropies, fractal dimension (FD), RMS, Hjorth parameters, Hurst exponent, KLC, LZC, and QRA. Detailed descriptions of the feature-extraction methods are provided in Table 4. To correctly detect focal and non-focal EEG signals, LS-SVM was adapted with feature ranking based on student t-test. To demonstrate the efficacy of features, some authors have investigated the several kinds of classifiers, including adaptive neuro fuzzy inference system (ANFIS), probabilistic neural network (PNN), naïve bayes (NB), non-nested generalized exemplars classifier (NNge), best first decision tree (BFDT), and radial basis function kernel-SVM (RBF-SVM). Though most of the research performed since has used the Bern-Barcelona dataset to design computer-aided systems capable of detecting epileptic seizure focus, some have explored different feature-extractions and machine-learning methods. Algorithmic progress has been limited due to the lack of a standard multi-channel benchmark iEEG datasets. Nevertheless, how the implementation of any of these methods will influence epileptic focus detection for real clinical applications is not yet clear.

Recently neurology hospitals measure ECoG with a high sampling frequency (20 kHz) and high-density electrode grids (multi-channels) in order to localize the epileptic seizure focus. Most recently, the connectivity and synchronization processes of the epileptic brain have been analyzed in different studies (Klimes et al. 2015; Warren et al. 2010; Antony et al. 2013). In general, the epileptic brain is distinguished by increased neuronal synchrony, and the epileptic focus is often functionally disconnected from the surrounding areas (Warren et al. 2010; Antony et al. 2013). The energy variation of the SOZ is significantly higher than that of the non-epileptic zone of the ripple (80–250 Hz), and fast ripple (250–600 Hz) bands (Klimes et al. 2015; Warren et al. 2010; Antony et al. 2013). In particular, the repetitive waveform pattern called the high-frequency oscillations (HFOs) consists of ripple, and fast ripple are the valid frequency bands to localize the epileptic seizure focus (Jacobs et al. 2009; Zijlmans et al. 2011; Urrestarazu et al. 2007). Based on these recent findings related to epileptic focus and clinical application, the computer-aided solutions that utilize feature-extraction methods encompass several open research questions that deserve to be addressed. The Bern-Barcelona dataset does not support the neurological and clinical biomarker evidence that is collected through use of ripple (80–250 Hz) and fast ripple (250–600 Hz) bands.

In a clinical situation, the number of SOZ and non-SOZ channels in the cortex area depends on the patient, which creates an imbalanced problem. In our survey, only two recent studies related to the design of a computer-aided solution, those proposed by Akter et al. (Akter et al. 2020a, b) addressed these unresolved questions. In their first study, the information-theoretic entropy features were investigated in order to design a computer-aided solution for identification of the SOZ channels of patients with drug-resistant epilepsy. The same research team recently observed the statistical features of high-frequency bands of interictal iEEG that efficiently identify the epileptic focus channels. To assess the possibility of building a patient-independent design (PID), they compared their proposed patient-dependent system with a patient-independent system for localizing SOZ channels through use of the optimal classifier LGBM. Ultimately, their system to extract features from high-frequency iEEG signals and thereby detect epileptic focus completely satisfies neurological hospital measurement standards and complements recent research findings related to epileptic seizure focus. However, many investigations are still required to improve the system’s performance and usability (Akter et al. 2020b).

With the rapid growth of neural-network-based methodologies in recent decades, the computer-aided designs that use these methods have been introduced to assist clinical experts in the diagnosis and treatment of epilepsy.

The evaluation criteria are the key factors that measure the efficacy of the computer-aided solutions. For completeness, a brief review of the concepts of evaluation of performance are presented to know what matrices must be considered during evaluation computer-aided solutions to the problem of epileptic focus localization.

This survey aimed to provide a comprehensive overview of the current epilepsy research focused on developing a computer-aided system for the localization of epileptic focus detection. We have surveyed the different biomarkers- and feature-extraction-based approaches developed between 2013 and 2020 to facilitate creation of an AI-based diagnosis-aid. The numerous explored approaches can be divided into four main categories: (i) HFOs, (ii) PAC, (iii) IED, and (iv) feature-extraction methods.

Some epilepsy studies with a variant of public datasets, including Bonn, Flint Hill, Freiburg, CHB-MIT, Epilepsia, and TUSZ, introduced computer-aided solutions to identify epilepsy patients or seizure patterns in EEG signals. For instance, the Bonn dataset consists of five subjects with single-channel surface EEG at a sampling rate of 173.61 Hz, denoted as Set A (healthy awake and eyes open), Set B (healthy awake and eyes closed), Set C (Epileptic seizure-free interictal), Set D (Epileptic seizure-free interictal), and Set E (epileptic seizure activity). A two-class problem was formed by assigning class labels to epileptic activity (set A) vs others (Set B, C, D, and E). Therefore, different combinations, such as E vs. A, E vs. B, E vs. C, E vs. D, and so on, were performed to detect epileptic vs. non-epileptic activities. However, this dataset’s main goal was to detect epilepsy patients from other subjects. Similarly, the other datasets listed in Table 2 only focused on detecting either seizure or non-seizure patterns in an EEG segment.

For drug-resistant epilepsy, the primary clinical problem is identifying the epileptic focus that must be removed entirely in order to halt seizures. To identify the epileptic focus for surgery, recent neurology hospitals have used grid electrodes (multi-channel) with a high sampling rate to observe ripple (80–250 Hz) and fast ripple (250–600 Hz) bands. The underlying idea behind using a high sampling rate is that ripple and fast ripple bands bear the potential guide for surgical treatment of drug-resistant focal epilepsy. Of note, the epilepsy biomarkers (HFOs, PAC, and IEDs) may occur during ictal, preictal, and interictal states, and the rate of biomarker events tends to be higher in the SOZ channels. As discussed in the biomarker section, many research works based on AI aim to develop automated methods for finding these biomarkers in multi-channel EEG signals. However, the biomarkers in EEGs are differentiated into ripples and fast ripples in EEG signals, and the identification of the biomarkers requires separate ripple and fast ripple iEEGs to identify SOZ channels. When designing a computer-aided solution using these biomarkers, most of the algorithms encounter a common problem: precisely calculating baseline activity. This is because the background activity is not perfectly flat in multi-channel EEG. It is notable that there are several possible subgroups of neurological biomarkers in EEG signals, including HFOs (ripples, fast ripples, and ripples+fast ripples) and IEDs (spikes, wave complex, polyspike, slow-wave, and artifacts). These biomarkers’ variants collectively hint at an intricate AI solution for the identification of possible SOZ channels. However, due to the lack of a standard public dataset related to the biomarkers and their distinctive performance evaluation, it is infeasible to compare the literature’s biomarker-related AI systems.

According to the various studies in this paper, many techniques have introduced the feature-extraction methods described in Table 5 to identify epileptic focus from EEG signals collected in public and hospital datasets. The studies listed in Table 5 collectively provide evidence that the information-theoretic and statistical features are another engineering marker that may help researchers to identify the epileptic focus. Among the many significant advantages of designing a computer-aided system is that it does not need to find a variant of biomarkers in EEG signals, as the statistical features measure the data dispersion from EEG signals to characterize normal and epileptic activities. It is important to note that EEG signals recorded from epilepsy patients may consist of epileptic biomarkers during ictal and interictal phases. The use of statistical methods may provide exciting tools that allow scholars to isolate the normal and epileptic properties from EEG signals and eventually design computer-aided solutions for the identification of SOZ channels.

Based on our survey, we also wish to identify some research problems that must be resolved before more efficient computer-aided solutions can be designed following:To conclude, this review suggests that the development of robust, efficient, and convenient computer-aided systems based on feature-extraction methods for identifying SOZ channels provides a new direction in epilepsy research. The development of a trustworthy system remains an open challenge. The next generation of machine-learning approaches to computer-aided diagnosis systems will have to take the user into account. Furthermore, we would like to emphasize that the collaboration between engineering and medical sectors is the most crucial key to the success of a practical computer-aided diagnosis aid.