AI-Based Electroencephalogram Analysis in Rodent Models of Epilepsy: A Systematic Review

Abstract

About 70 million people globally have been diagnosed with epilepsy. Electroencephalogram (EEG) devices are the primary method for identifying and monitoring seizures. The use of EEG expands the preclinical research involving the long-term recording of neuro-activities in rodent models of epilepsy targeted towards the efficient testing of prospective antiseizure medications. Typically, trained epileptologists visually analyse long-term EEG recordings, which is time-consuming and subject to expert variability. Automated epileptiform discharge detection using machine learning or deep learning methods is an effective approach to tackling these challenges. This systematic review examined and summarised the last 30 years of research on detecting epileptiform discharge in rodent models of epilepsy using machine learning and deep learning methods. A comprehensive literature search was conducted on two databases, PubMed and Google Scholar. Following the PRISMA protocol, the 3021 retrieved articles were filtered to 21 based on inclusion and exclusion criteria. An additional article was obtained through the reference list. Hence, 22 articles were selected for critical analysis in this review. These articles revealed the seizure type, features and feature engineering, machine learning and deep learning methods, training methodologies, evaluation metrics so far explored, and models deployed for real-world validation. Although these studies have advanced the field of epilepsy research, the majority of the models are experimental. Further studies are required to fill in the identified gaps and expedite preclinical research in epilepsy, ultimately leading to translational research.

1. Introduction

Epilepsy is a heterogeneous neurological disease characterised by recurrent seizures [1,2]. Although seizures are the primary hallmark of epilepsy, the clinical scenario is sometimes made more complex by the existence of comorbidities [3]. The physiologic feature of seizures is the transient excessive and synchronous discharges by a group of neurons in the brain, leading to the behavioural alteration of a person or animal [4]. Such behavioural alteration includes sensory, autonomic, cognitive, and emotional [5]. Nevertheless, these changes are not specific to epileptic seizures, and not all seizures are epileptic; syncope, psychogenic non-epileptic seizures, migraine, parasomnias, transitory ischaemic attack, paroxysmal dystonia, and non-epileptic myoclonus are some of the conditions in adults that imitate epileptic seizure [6,7]. The identification of epileptic seizure semiology from other non-epileptic conditions is crucial in the diagnosis of epilepsy. According to the International League Against Epilepsy (ILAE), epilepsy is defined by any of the following conditions: (i) At least two unprovoked (or reflex) seizures occurring >24 h apart. (ii) One unprovoked (or reflex) seizure and a probability of further seizures similar to the general recurrence risk (at least 60%) after two unprovoked seizures occurring over the next ten years. (iii) Diagnosis of an epilepsy syndrome [1].

An estimated 70 million people globally have been diagnosed with epilepsy [8], which is a global health burden with adverse social, economic, and psychological impacts on people with epilepsy [9,10]. The occurrence and frequency of epilepsy vary between countries, with an estimated incidence rate of 49 cases per 100,000 individuals in high-income countries and 139 cases per 100,000 individuals in low- and middle-income countries (WHO, 2023). The estimated prevalence rate of active epilepsy is 6.38 per 1000 individuals [11] and 4 to 10 per 1000 individuals (WHO, 2023).

The epilepsy research community has devoted decades of time and resources to studying epilepsy and researching antiseizure drugs (ASDs) in an effort to mitigate the symptoms of epilepsy. Despite the number of ASDs available, one-third of people with epilepsy experience resistance to pharmacotherapy [12], where ASDs fail to control seizures. This group of people bear the major burden of epilepsy in the general population. Thus, a great unmet clinical need exists for identifying and developing new ASDs that efficiently manage pharmaco-resistant spontaneous recurrent seizures (SRSs).

The complexity of epilepsy and pharmaco-resistant seizures cannot be thoroughly researched in the clinical setting with humans for ethical reasons [13]. Although cellular models are frequently employed initially to clarify molecular pathways in disease processes, animal models are increasingly useful for preclinical research in studying human diseases and clinical aspects [14,15] and discovering new drugs and drug targets [16]. In the study of epilepsy and its resistance to ASDs, no singular animal can model the different types of epilepsy [17]. Most animals with central nervous systems are likely to experience epilepsy, and dogs, cats, primates, rats, mice, and zebrafish have all been used to model different types of epilepsy [18]. The selection of a particular animal may be based on practical needs [18]. Although humans and rodents such as mice have different neurobiologies, mice are frequently used to model human neurological disorders [19,20], and recently, rodents (rats and mice) have been prioritised for epilepsy research [18].

Electroencephalography (EEG), introduced by Hans Berger, is a clinical procedure for reading spontaneous neural activities [21]. The ability to read the electrical signals in the brain can provide insight into the abnormalities that may occur in the brain. It may be useful in assisting with the diagnosis or assessing the presence of comorbidities in various neurological disorders. With the use of EEG, neurologists can understand the alterations in the human brain that accompany epileptic seizures. Its usage extends to rodent models of epilepsy. The analysis of the rodent brain fluctuations in rodent models of epilepsy uncovers the disease development, leading to understanding disease mechanisms and evaluating the effects of ASDs and experimental treatments.

The waveforms in EEG activity that characterise seizures are known as epileptiform discharges [22]. Epileptiform discharge is a transient burst of spikes, polyspikes, and polyspike–wave and spike–wave complexes with varying shapes and amplitudes indicative of cortical hyperexcitability and disruption [22,23]. Generally, epileptiform discharge is categorised into four states: interictal, preictal, ictal, and postictal.

• Interictal discharge: Interictal spikes are the intermittent neurological discharge observed between seizures in patients and animals with epilepsy [24,25]. Interictal spikes are noncoincidental [26] but relatable to seizures, and they are quantified as the most reliable epileptogenic biomarker [27,28].
• Preictal discharge: The preictal state is the abnormal neural discharge just before the manifestation of a seizure event, where a patient may feel the presence of seizure aura, which are physiological changes such as muscle twitches and gastrointestinal upset [29]. The temporal and spatial changes of the preictal discharge correlate with seizure onset; towards the onset of seizure, the amplitude of preictal discharge increases, while the interval decreases [30]. Thus, understanding and detecting the preictal spike mechanism can predict the next seizure episode [31].
• Ictal discharge: This is the most symptomatic and shortest phase of the epileptiform discharge. The ictal spike represents the critical event of a seizure that characterises an active epileptic condition.
• Postictal discharge: This state of the brain describes the abnormal neurological performance that begins at the end of a seizure episode; this state may last for hours before the resumption of the neural baseline activity [32]. Postictal discharge receives less attention than the interictal and preictal states. Nevertheless, the postictal characteristics may assist in distinguishing epileptic and non-epileptic seizures [33]

Epileptiform discharge identification includes seizure detection (the identification of ictal discharges), seizure prediction (the identification of preictal discharges), and seizure type classification (the categorisation of the different types of seizures). Increasingly, epilepsy research is moving towards determining the disease-modifying effects of drugs, which requires persistent assessment of brain activities, including epileptiform discharges, through long-term recording of EEG in animal models of epilepsy [34]. Traditionally, detecting epileptiform discharges can be achieved through visual analysis of the long-term EEG signal by a team of epileptologists. However, there are certain drawbacks to the visual identification of epileptiform discharges in long-term EEG recordings. In addition to the time spent reviewing the EEG recording, the subjective seizure identification between epileptologists on the same EEG recording due to various seizure morphologies and the similarity of seizure patterns with noise and artefacts is another major challenge.

In contrast to the manual detection of seizures, algorithmic approaches have been explored to analyse EEG signals automatically. Researchers are moving towards improving clinical practice by applying machine learning (ML) to analyse long EEG recordings. Therefore, this study aims to present a systematic literature review to identify the processes and state-of-the-art ML and deep learning (DL) detection of epileptiform discharges in rodent models of epilepsy.

The remainder of this paper is structured as follows: Section 2 presents the classification of epilepsy, background on rodent models of epilepsy, EEG signals, and ML and DL techniques; Section 3 describes the research methodology and research questions adopted for the organisation and execution of this systematic literature review; Section 4 presents the findings from the surveyed literature; Section 5 discusses the implications of our findings; and the Conclusion is presented in Section 6.

2. Fundamental Concepts and Background

2.1. Classification of Epilepsy

The ILAE has updated the classification of epileptic seizures, taking into account their aetiology, symptoms, and underlying physiological mechanisms. The classification adopted by most neurologists and epileptologists was developed by a committee of the ILAE in 1981 [35]. The classification implements types of seizures as simple partial seizure, complex partial seizure, generalised tonic–clonic seizure, absence seizure, secondarily generalised tonic–clonic seizure, and others. The classification, though useful, was limited in classifying several focal motor seizures, such as tonic, clonic, atonic, or myoclonic [36]. Resolving the limitation of this earlier classification, ILAE 2017 introduces a three-step classification system of epilepsy (seizure type, epilepsy type, and epilepsy syndrome) [37] with emphasis on the aetiology and comorbidities at each step of the classification [38]. The starting point for classifying epilepsy is the determination of the seizure type. This implies the identification of the seizure onset, which could be focal, generalised, or unknown. Focal seizures happen when a small network of neurons within one lobe (hemisphere) is hyperexcited and can spread to neighbouring neural regions [39]. In contrast, generalised seizure involves neuronal hyperexcitability of all the hemispheres of the brain from the outset [37]. The types of epilepsy include focal epilepsy resulting from focal seizures, generalised epilepsy, which is characterised by generalised seizures, combined focal and generalised epilepsy, and epilepsy of unknown seizure onset. Seizures can begin in any of the brain lobes, but the most common is the mesial temporal lobe, which holds the amygdala and hippocampus [39,40]. Hence, the most common type of epilepsy is temporal lobe epilepsy (TLE) [13]. The final classification step, epilepsy syndrome, is defined by the ILAE as “a characteristic cluster of clinical and EEG features, often supported by specific etiological findings (structural, genetic, metabolic, immune, infectious, and unknown)” [38]. Syndromes frequently exhibit manifestations that vary depending on the age of the individual, and in certain instances, they may wane at specific ages [38].

2.2. Rodent Models of Epilepsy

The rodent models of epilepsy are categorised into genetic and induced (see Figure 1). In both categories of modelled epilepsy, seizures are the fundamental phenotypic feature of assessment. There are three primary ways of inducing epilepsy in animal models: chemo-convulsant (direct or systemic), electrical stimulation, and acoustic stimulation [41]. The most common is the use of chemo-convulsants to mimic epilepsy similar to a particular type of human epilepsy [42].

The systemic or direct injection of chemo-convulsants (kainic acid, pilocarpine, pentyle-netetrazol) into the rodent brain and electrical and acoustic brain stimulation induce brain injury [13,43], resulting in status epilepticus (SE) [41]. SE is a generalised seizure activity that lasts more than 5 min without full recovery [44]. The brain analysis of these animals following the administration of the chemo-convulsants reveals substantial damage in the olfactory cortex, amygdala, thalamus, neocortex, hippocampus and substantia nigra [45]. To reduce the severity of SE that could lead to permanent neuronal damage, the SE is abated with the injection of sedatives, leading to a latent phase before spontaneous recurrent seizure (SRS) begins to manifest days later [46]. Typically, rodents that have been subjected to SE go on to develop epileptiform discharges and chronic SRS [47]. During the chronic phase of SRS, either the pathophysiology of epilepsy is studied or new therapeutic approaches for suppressing SRS are tested [48].

Although epilepsy is defined clinically as unprovoked seizures from genetic and acquired factors, the studies on epilepsy have traditionally relied on induced acute seizure models. While induced acute seizure models such as maximal electroshock seizure and pentylenetetrazole are invaluable for initial screening of anticonvulsant activity, it is important to note that these models do not fully represent the complex pathophysiology of epilepsy [49]. These models are used primarily for their efficiency in identifying potential ASDs. Advances in genomics have identified genes linked to epilepsy [49]. In genetic models of epilepsy, clinically relevant mutations are identified and replicated in mice, known as knockout or knock-in mutagenesis [41]. These genetic models provide essential insight regarding the role of mutation in ictogenesis and epileptogenesis [50]. Genetic models such as audiogenic seizure-susceptible DBA/2 mouse is a relevant model for studying sudden death in epilepsy [51] due to its susceptibility to audio generalised seizures evoked by excessive auditory stimulation [52]. The genetic absence seizure rats (GAERS) model in Wistar rats represents a prototypical form of childhood generalised idiopathic epilepsy [53]. Dravet syndrome is an intractable epileptic encephalopathy [54]. Almost 90% of Dravet syndrome patients have de novo heterozygous mutation in the SCN1A gene [55]. Replicating the causal mechanism of Dravet syndrome relies on generating animals with a genetic mutation in the SCN1A gene [49]. Like humans, the SCN1A is expressed throughout the central nervous system of rodents [56].

2.3. EEG Signal

There are two types of EEG recording: noninvasive and invasive. Noninvasive EEG has electrodes placed on the surface of the scalp. While noninvasive EEG is widely used for monitoring electrical activities in the brain, it is important to acknowledge its limitations. Surface EEG, in particular, is subject to inter- and intra-individual variability, which can result from factors such as electrode placement and individual physiological differences. Additionally, EEG’s spatial resolution is limited, making it difficult to precisely localize neural activity, especially in deeper brain regions. These limitations should be considered when interpreting EEG findings and underscore the need for complementary imaging modalities to provide more comprehensive spatial information. The invasive variation is most commonly used to diagnose patients with refractory epilepsy [57,58] and on animals in neurological research [59,60]. Invasive EEG recording can be achieved in two ways: (1) by placing the electrode on the exposed cortex of the brain under the dura, known as electrocorticogram (ECoG), and (2) by intracranial EEG (iEEG), in which the electrodes are inserted into the brain parenchyma [61,62]. EEG scalp electrodes cannot effectively detect seizure onset from the mesial temporal, orbitofrontal, or inter-hemispheric regions because of the enfolding of two-thirds of the cortex in sulci [63]. Following the proximity to the brain, the signals measured through invasion have better resolution with higher amplitude [64] and less susceptibility to biological artefacts (eye movements, muscle activities, and heartbeats) [62,65] compared to the scalp electrodes. The advantages of invasive versus noninvasive include the following: spatial resolution (tenths of millimetres versus centimetres), broader bandwidth (0–500 Hz versus 0–50 Hz), and higher characteristic amplitude (50–100 μV versus 10–20 μV) [66].

Preclinical procedures involving the use of animals and their care must obtain ethical approval and conform to institutional and international animal protection and care guidelines. Following the ethical approval, Wang et al. [67] defined two procedural steps (electrode preparation and surgical implantation of invasive electrodes into the skull) before the inducement of seizures in rodent models of epilepsy or the EEG signal waveform recording of the genetic models of epilepsy.

The implanted electrodes pick up the nearby neural activities to generate EEG signals; the EEG signal undergoes amplification, filtering, and digitisation before the transmission through wired or wireless telemetry [68]. The EEG signal from the different EEG recording electrodes (also known as channels) is characterised according to amplitude, frequency, morphology, synchrony, symmetry, continuity (rhythmic, intermittent, or continuous), location, and reactivity [69]. However, frequency, amplitude, and morphology are the most used pattern-recognition methods for classifying EEG waveforms [70]. Inherent in the EEG signals is different information describing the different states of the brain through the frequency domain [71].

Human brain waves measured through noninvasive EEG are mostly composed of low-frequency oscillation (delta, theta, alpha, beta, and gamma) in the range of 0–100 Hz [72]. Additional waveforms ensue when brain activity is calibrated using invasive EEG [72]. These waveforms are high-frequency oscillations (HFOs), are categorised as high gamma, ripple, and fast ripple, and have frequencies between 100 and 500 Hz [73,74]. The HFOs are unconventional EEG frequencies discovered to be clinically significant through digital signal processing [69]. However, HFOs are often obscured by high-frequency artefacts; analysing these frequencies requires great care [73]. Furthermore, amplitude measures waveform voltage (magnitude). The frequency and amplitude of physiological waveforms typically have an inverse relationship; waves with lower frequencies (delta, theta) have larger amplitudes, while waves with higher frequencies (alpha, beta) have lower amplitudes [75]. Amplitude is an essential descriptor for detecting interictal abnormalities in the EEG signals [76,77,78]. Morphology describes the overall shape and pattern of EEG transient events (spikes, sharp waves, and spike-and-wave complexes), which are distinct from the background activity and are associated with neurological disorders [79].

2.4. Machine Learning and Deep Learning Techniques

ML and DL are artificial intelligence (AI) branches implementing mathematical models and algorithms [80]. The ML algorithms receive historical data as inputs, learn from the data, and make decisions without explicit programming. DL is also a data-driven learning technique where computation is carried out through multi-layer neural networks. With AI, it has been possible to analyse and discover patterns not obvious to the human eye from vast amounts of data.

Developing ML and DL models involves a series of steps exemplified in Figure 2. Preprocessing the signals, or signal processing, is a mandatory phase in the ML and DL frameworks after acquiring the raw EEG signal. EEG signals have a high temporal resolution, making them susceptible to noise and artefacts [81]. Extrinsic and intrinsic (physiological) factors cause artefacts. Extrinsic factors such as faulty electrodes, line noise (50 Hz or 60 Hz interference for European and North American standards, respectively), and high impedance can be avoided by adhering to precise and strict recording procedures [82]. Physiological artefacts, which include eye movements, eye blinks, and cardiac and muscle activities, are more challenging to avoid [83]. The interference of these artefacts with neural information may render misleading inferences when used in practice to study varying neural disorders [84,85]. Following the proximity to the brain, the signals measured through invasive methods have better resolution with higher amplitude [64] and less susceptibility to physiological artefacts [62,65]. However, the signal processing phase is primarily concerned with attenuating artefacts that may skew the signal content [86] and, sometimes, normalising data into a uniform range. The choice of preprocessing method is based on the preferred characteristics of the signal and the artefacts of concern. Generally, most studies use bandpass filters to filter out the required signal frequency.

The feature engineering phase comprises feature extraction and dimension reduction. In the feature extraction stage, preprocessed signals are transformed into relevant attributes of the signals that preserve the physiological activities in the preprocessed signals [87]. EEG signal feature extraction can be performed in the frequency, time, or time–frequency domains [88]. Time domain (TD) analysis is computed on the raw EEG signal involving the amplitude over time. Unlike other domains, TD does not require a feature extraction tool. In contrast, the frequency domain (FD) analysis provides an intuitive grasp of the frequency or spectral components of the EEG signal and is computed on the discrete Fourier transform. Short-time Fourier transform (STFT) spectrograms and discrete wavelet transform (DWT) are commonly applied to decompose signals in the time–frequency domain (TFD) to study the changes in frequency over time [88]. Regardless of the domain, EEG signal features are generally categorised into linear and nonlinear features [89]. The brain is a chaotic, dynamic biological system. The generated EEG signals produce amplitude that varies arbitrarily with respect to time. The underlying neural subsystems that produce the EEG signals are generally considered nonlinear [90]. After feature extraction, the curse of dimension is a critical issue when applying machine learning algorithms to high-dimensional data [91]. Dimension reduction is a technique to reduce high-dimensional features in a dataset into a low-dimensional space while preserving the original information; this process can be achieved through feature selection or feature transformation. Feature selection involves selecting a subset of discriminative features from the universal set of features extracted [92,93]. In contrast, feature transformation converts the initial extracted features to produce new features that more accurately capture the underlying relationships or patterns in the data. Feature engineering plays a fundamental role in classical ML processes, as the performance of the ML model depends on the quality of the extracted features [94]. In Figure 2, a DL model follows the same processing steps as ML modelling. However, feature engineering is automatic as opposed to manual feature engineering in ML modelling.

Figure 2. The methodological approach to the development of ML and DL models: Following the acquisition and preprocessing of the EEG signals, both ML and DL commonly involve the phases of classification and performance evaluation. In ML modelling, feature extraction is performed explicitly and requires manual engineering. However, in DL modelling, feature extraction is automatic [95].

Figure 2. The methodological approach to the development of ML and DL models: Following the acquisition and preprocessing of the EEG signals, both ML and DL commonly involve the phases of classification and performance evaluation. In ML modelling, feature extraction is performed explicitly and requires manual engineering. However, in DL modelling, feature extraction is automatic [95].

3. Materials and Methods

This review aims to evaluate the literature on the application of ML and DL in analysing EEG epileptiform discharges in rodent models of epilepsy. This review was systematically performed in conformity with the preferred reporting items for systematic reviews and meta-analysis (PRISMA) guidelines [96]. The systematic protocol described below is registered on the International Platform of Registered Systematic Review and Meta-analysis Protocol (INPLASY) with the DOI number: 10.37766/inplasy2024.7.0108:

• Defining the research questions;
• Execution of article searches within specified databases;
• Filtering articles through the evaluation of their relevance;
• Data extraction;
• Synthesizing of results.

3.1. Research Questions

The following research questions are answered in this review study:

• RQ1: What rodent models of epilepsy and seizure/epilepsy types have been automatically analysed with ML or DL algorithms?
• RQ2: What features and feature engineering techniques have been considered in the classical machine learning detection and prediction of seizures in the rodent model of epilepsy?
• RQ3: What ML or DL methods have been exploited in detecting and predicting seizures from EEG of rodent models of epilepsy?
• RQ4: What training methodologies and evaluation metrics have been used in the rodent models of epilepsy, and which of the developed DL/ML models have been implemented?

3.2. Search Execution

Selecting precise search terms to retrieve broad and relevant articles was challenging. We combined different search terms and their synonyms into logical search strings (S1–S4). After preliminary searches yielded few articles, we focused on articles published in the last three decades, between 1 January 1994 and 1 January 2024. Articles not retrieved by the S1 and S2 search strings were retrieved by quoting the search phrases in the S3 and S4 logical search strings. These logical search strings were used to search the Google Scholar [97,98] and PubMed databases [99]:

• S1: ((EEG OR Electroencephalogram) AND (Seizure detection OR Machine learning seizure detection OR Deep learning seizure detection OR Interictal spike detection OR Ictal spike detection) AND (Rodent model of epilepsy OR Mouse model of epilepsy OR Rat model of epilepsy)).
• S2: ((EEG OR Electroencephalogram) AND (Seizure prediction OR Machine learning seizure prediction OR Deep learning seizure prediction OR Spike detection OR Preictal spike detection) AND (Rodent model of epilepsy OR Mouse model of epilepsy OR Rat model of epilepsy)).
• S3: ((“EEG” OR “Electroencephalogram” OR “iEEG”) AND (“Seizure detection” OR “Machine learning seizure detection” OR “Deep learning seizure detection” OR “Interictal spike detection” OR “Ictal spike detection”) AND (“Rodent model of epilepsy” OR “Mouse model of epilepsy” OR “Rat model of epilepsy”)).
• S4: ((“EEG” OR “Electroencephalogram” OR “iEEG”) AND (“Seizure prediction” OR “Machine learning seizure prediction” OR “Deep learning seizure prediction” OR “Spike detection” OR “Preictal spike detection”) AND (“Rodent model of epilepsy” OR “Mouse model of epilepsy” OR “Rat model of epilepsy”)).

A total of 3021 articles were retrieved through the four search strings. Precisely, Google Scholar returned 2443 articles, while PubMed returned 578. According to PRISMA protocol, the procedure for the literature search, document selection, and filtering is summarised in Figure 3.

3.3. Article Selection and Filtering

A total of 3021 articles were retrieved according to the search procedure. After the removal of 1012 duplicate articles, 2009 articles were left for the eligibility screening. Following the screening of articles for eligibility using the inclusion and exclusion criteria and evaluation of abstracts, 1965 articles were excluded (articles not written in English: 3, other exclusion criteria: 1962), leaving behind 44 articles for further assessment. The abstract evaluation specifically targeted the identification of articles that utilised computational analysis of EEG data collected from rodent models of epilepsy. After reading the full text, 23 studies were discovered to have used nonmachine learning and deep learning techniques. These studies were excluded, and 21 articles based on ML/DL techniques met the eligibility criteria. One article from the reference list was also found to be relevant. The articles were independently filtered by two researchers. A total of 22 papers were considered eligible and included in this review study.

• Inclusion criteria:
  • Conference papers published by ACM, IEEE, or Springer;
  • Published between 1 January 1994 and 1 January 2024;
  • Focused on automatic seizure detection and prediction with machine learning and deep learning techniques;
  • The study is conducted using EEG on rodent models of epilepsy;
  • Full text available;
  • The study reported in the article is empirical.
• Exclusion criteria:
  • The study uses EEG from humans or other animals, not rodents;
  • Full text is inaccessible;
  • Literature review studies, books, short papers/abstracts;
  • The study is focused on other neurological disorders (Alzheimer’s, dementia, sleep disorder, etc.);
  • The study uses other clinical methods of determining epilepsy (e.g., genomics);
  • Studies not written in English;
  • Studies focusing on clinical studies only;
  • Studies performed using other nonmachine or deep learning techniques;
  • Journals not listed in Journal Citation Report (JCR).
• Data Extraction
  Relevant data were gathered and analysed to synthesise and answer the research questions from the included studies. In the process, the following data fields were extracted:

  –

  D1: The aim of the article.

  –

  D2: The rodent models of epilepsy analysed.

  –

  D3: The epileptiform discharge recognition task: detection or prediction of seizures.

  –

  D4: The feature engineering techniques adopted.

  –

  D5: Employed ML or DL technique, either classical or a combination of techniques and evaluation result.

  –

  D6: The implementation of the ML or DL models.

4. Results

4.1. Ml and DL Article Distributions

Figure 4 shows the number of yearly published articles that employed ML and DL techniques. The usage of these techniques started with classical ML in 2006 and continued with a steady frequency until 2011. DL emerged in rodent EEG seizure detection and prediction in 2019. However, it is noticeable that the frequency of DL articles in rodent EEG analysis has not surged since its first application.

4.2. Distribution of Articles among Epileptiform Discharge Tasks

Among the eligible articles, only one article focused on seizure prediction. The 21 other articles focused on seizure detection. Overall, there has not been a growing number of articles published regarding the detection and prediction of seizures in rodent models of epilepsy, as shown in Figure 5.

4.3. RQ 1: What Rodent Models of Epilepsy and Seizure/Epilepsy Types Have Been Automatically Analysed with ML or DL Algorithms?

Different seizure types have been induced through various rodent models of epilepsy. Several previous works [34,100,101,102,103,104] evaluated temporal lobe epilepsy, and eight studies focused on absence seizures in genetic absence epilepsy rats [103,105,106,107,108,109,110,111]. Two studies analysed Dravet syndrome [34,112] and tonic–clonic seizures [109,113], respectively. Convulsive seizures [114], clonic seizures [115], frontal lobe epilepsy [116], and post-traumatic epilepsy [117] were each analysed by one study, respectively. The majority of the studies analysed a seizure type, nine studies focused on seizures of generalised onset (absence and generalised tonic–clonic seizures), while seven studies focused on detecting seizures in focal onset seizure types. Among the studies reviewed, three investigated two seizure types [34,103,109], whereas others [118,119,120] did not specify the type of seizure analyzed. Refer to

Table 1. Summary of rodent models and seizure types analysed in recent studies.

Ref	Rodent Model	Seizure/Epilepsy Type Modelled
[110]	$\gamma$-aminobutyric acid A (GABAA) receptor mutation ($\gamma$2R43Q)	Absence epilepsy
[102]	Intrahippocampal kainic acid	Mesial temporal lobe epilepsy
[112]	High temperature	Dravet syndrome
[109]	Genetic and kainate systemic injection	Absence and tonic–clonic seizures
[103]	Genetic and kainate systemic injection	Absence and limbic seizures
[108]	Genetic	Absence seizures
[107]	Genetic	Absence seizures
[120]	-	-
[34]	Intra-amygdala kainic acid, Dravet, Pilocarpine	Temporal lope epilepsy, Dravet syndrome
[115]	Pentylenetetrazole	Clonic seizures
[106]	GAER	Absence seizures
[104]	Perforant pathway stimulation	Mesial temporal lobe epilepsy with hippocampal sclerosis
[114]	Pilocarpine hydrochloride	Convulsive seizures
[113]	Pilocarpine injection	Generalized tonic–clonic seizures
[105]	GAERS	Absence seizures
[117]	Perforant path kindling susceptible	Post-traumatic epilepsy
[116]	Kindling	Frontal lobe epilepsy

for a comprehensive list of seizure types and the corresponding rodent models of epilepsy.

4.4. RQ2: What Features and Feature Engineering Techniques Have Been Considered in the Classical Machine Learning Detection and Prediction of Seizures in Rodent Models of Epilepsy?

Studies have investigated different features and feature extraction techniques. Pan et al. [111] analysed four linear time-domain features, mean value, curve length, accumulated energy, and sixth power, and one nonlinear dynamics feature, average nonlinear energy. Contrary to extracting features, in an extension study, Pan et al. [118] used the weighted locally linear embedding (WLLE) and locally linear embedding (LLE) in reducing the dimension of the EEG time-domain signal. Shin et al. [100] explored the three domains of EEG signal and extracted ten features; one FD feature (periodic frequency) was extracted using the fast Fourier transform, while Daubechies wavelet decomposed the time-domain to acquire detailed and approximate coefficients from wavelet levels 1–4 and one TD feature (mean of the absolute amplitude). In addition to the mean energy and mean curve length, wavelet transform was used to extract the wavelet energy in the beta band (12–20 Hz) by Nandan et al. [119]. Wang et al. [101] used digital filters to exclude and study the spectral power of ten frequency bands, together with approximate entropy, a nonlinear feature that measures the predictability of a time series. Fumeaux et al. [102] extracted 141 linear features from TD and FD. The high-dimensional data were reduced using principal component analysis (PCA). Studies based on deep learning investigate the signals directly in the time and frequency domains without the need to extract features from these domains [104,105,113,114,117].

Table 2. Previous work on feature types and feature engineering techniques.

Author	EEG Domain	Feature Extraction	Feature Type	Feature Selection	Feature Transformation
[110]	FD	Wavelet	12 linear	-	-
[102]	TD, FD	-	141 linear	-	PCA
[112]	TD, FD	Gabor transform	9 linear	Neighborhood component analysis	-
[109]	FD	Wavelet	12 linear	Forward feature selection	-
[103]	FD	Wavelet	2 linear	Forward feature selection	-
[108]	TD, FD	FFT	2 linear, 1 nonlinear	-	-
[107]	FD	FFT	1 linear, 1 nonlinear	-	PCA
[120]	TD	PCA, Laplacian Eigenmap manifold	linear	-	PCA and Laplacian Eigenmap manifold
[34]	TD, FD	Wavelet	19 linear
[115]	TD	Natural visibility algorithm, Markov–binary visibility graph,	14 linear	Sequential forward selection	-
[106]	FD	FFT	2 linear	-	-
[116]	FD, TFD	FFT, wavelet transform	1 linear, 1 nonlinear	-	-

details the feature extraction techniques and the type of EEG features adopted in the ML detection of seizures in rodent models of epilepsy. Time–frequency domain features were used in one study [116]. Most articles extracted their features from either TD or FD or both. The linear features have been intensively used compared to the nonlinear features. Three out of the total number of articles reviewed considered at least one nonlinear feature; [107,108] measured complexity with approximate entropy while [111] included nonlinear energy. Forward feature selection otherwise known as sequential forward selection, a type of sequential feature selection, was employed by three studies [103,109,115]. Three studies transformed features into a low-dimensional feature space using PCA [102,107,120]. In addition to using PCA, Ramirez et al. [120] used a Laplacian eigenmap manifold.

4.5. RQ3: What ML or DL Methods Have Been Exploited in Detecting and Predicting Seizures from EEG of Rodent Models of Epilepsy?

The thresholding technique is the earliest approach to the automatic detection of epileptiform discharge [121]. In this approach, a feature is computed for each consecutive window of an EEG signal, and the seizure segment is established by comparing the extracted feature with a defined threshold [122]. However, the lack of a gold standard threshold is a drawback as this approach highly depends on the value and type of threshold defined [34]. The benefit of machine learning over the threshold approach lies in its ability to learn and discover patterns inherent in the signal.

Consequently, Pan et al. [111] applied a support vector machine (SVM) and back-propagation neural network in differentiating the baseline signals from the seizure segments in the EEG of five rats. The SVM model outperformed the neural network model with an accuracy of 100%. However, the baseline and the seizure classes were highly imbalanced, with 38 epochs of baseline and 152 epochs of seizures. Pan et al. [118] investigated the performance of WLLE and LLE in combination with SVM in 16 male Swiss mice. WLLE separated the data into clusters, whereas LLE clusters overlapped. WLLE and SVM performed with an accuracy of 72.79–90.66% and a false-alarm rate of 0.00%. Shin et al. [100] assessed the performance of five- and ten-fold cross-validation in discriminating interictal and ictal spikes in five rats. The model trained with ten-fold cross-validation was evaluated to a considerable sensitivity of 93.03% and specificity of 94.97%. However, the lack of evaluation of the model with an independent test set and the imbalance of the 3106 interictal and 2356 ictal class imbalances places a limitation. To determine the type of SVM suitable for identifying ictal data, Nandan et al. [119] proposed using support vector data description (SVDD) in identifying interictal and ictal spikes. The model proffered 100% sensitivity and 79–85% specificity after training with two-fold cross-validation. Wang et al. [101] developed a linear discriminant-based model with the EEG signal from 11 rats. After two-fold cross-validation, the model evaluation was reported as 100% sensitivity, 22% false detection, and 1.69 s delay latency. Buteneer et al. [109] presented the recurrent neural network training with a reservoir-computing algorithm. The model was trained and tested with EEG signals from kainate-induced rats and genetic absence epilepsy rats (GAERS). After a ten-fold cross-validation, the GAERS presented a sensitivity of 96.4% and specificity of 96.2%. The kainate-induced rats had 93.9% sensitivity and 99.1% specificity. In an attempt to implement a RISC-like processor to detect and suppress epileptic seizure, Chen et al. [108] trained a linear least square algorithm with the first ten hours of EEG signal from two rats and tested on the remaining forty-eight hours of the same rats. Accuracies of 92.33% and 92.81% were achieved for rat1 and rat2, respectively. Rat1 and rat2 had a delay latency of 0.63 s and 0.62 s, respectively.

Table 3. Recent ML algorithms for detecting seizures in rodent models of epilepsy.

Author	Number of Rodents	Classifier	Result
			Accuracy (%)	Precision (%)	Sensitivity (%)	Specificity (%)	AUC (%)	False Alarm (%)	Delay Latency (s)
[110]	11	SVM	-	59.00	68.00	-	-	-	-
[102]	11	GLM	-	-	-	-	99.50	-	-
[112]	13	RF	92.80–97.20	77.30–97.40	93.30–95.80	91.90–98.90	-	-	-
[103]	45	RC-BRR	-	-	96.20	98.20	-	9.10	0.97
[107]	3	RBF-SVM	97.50	-	97.03	97.83	-	-	-
[120]	-	KRR	73.43–97.76	-	-	-	-	-	-
[34]	26	XGBoost	93.1–98.8	-	76.30–98.8	93.10–98.80	-	-	-
[115]	27	MLP	92.13	-	98.94	-	86.03	-	-
[106]	18	GNN	-	-	98.01	-	98.90	95.90	-

SVM = support vector machine; SVDD = support vector data description; RF = random forest; GLM = generalised linear model; LLS = linear least square; LDA = linear discriminant analysis; GNN = graph neural network; RC-RRN = reservoir computing with recurrent neural network; RC-BRR = reservoir computing with Bayesian relevance regression; KRR = kernel ridge regression; MLP = multi-layer perceptron.

summarises the recent studies on ML algorithms proposed for the detection of seizures in rodent models of epilepsy. Most articles used SVM compared to other algorithms applied in a single article. Impressive performance was achieved in recent articles from 2019 with the application of complex algorithms, namely, XGBoost, MLP, and GNN [34,106,115].

Machine learning-based seizure detection typically comprises feature extraction and classification stages. The tedious identification of appropriate features and the choice of classifier play pivotal roles in the performance of the ML seizure detection model [113]. The automatic extraction of features by the DL algorithm removes the hassle of feature extraction and poses a substantial advantage over classical machine learning [123]. Recently, DL has been investigated for seizure detection in rodent models of epilepsy.

Table 4. Recent DL algorithms for detecting seizures in rodent models of epilepsy.

Author	Number of Rodents	Classifier	Result
			Accuracy (%)	Precision (%)	Sensitivity (%)	Specificity (%)	AUC (%)	False Alarm (%)	Delay Latency (s)
[104]	7	Deep residual neural network	-	-	83.00	83.00	90.00	-	-
[114]	-	Sequential dual deep neural networks	-	98.00	100.00	-	-	6.00	-
[113]	32	Fully connected neural network	99.60–99.90	-	96.20–96.70	99.60–99.90	98.90–99.30	0.9–1.1	-
[105]	11	CNN	94.28	-	96.60	90.40	-	-	-
[117]	8	GoogleNet	81.10–95.60	-	-	-	-	-	-

summarises the DL techniques proposed for the detection of seizures in rodent models of epilepsy.

To date, only one study has focused on seizure prediction using frontal lobe epileptiform discharges. De et al. [116] examined the effectiveness of low-frequency wave ranges (0.5–13 Hz, commonly observed during sleep) and total frequency range (0.5–40 Hz) across four classifiers to predict seizures in WAG/Rij rats induced through kindling. The study did not state the number of rats in the study. However, five datasets were employed in the study. The classifiers examined included radial basis function–support vector machine (RBF-SVM), quadratic discriminant analysis (QDA), linear discriminant analysis (LDA), and linear support vector machine (LSVM). These classifiers were trained with wavelet coefficients and power spectral density features. The RBF-SVM classifier achieved the highest classification performance compared to all other classifiers across the five datasets. For the low-frequency signal, the maximum performance includes 94.62% accuracy, 98.46% sensitivity, and 90.76% specificity. The performance of the total frequency signals was 96.15% accuracy, 100% sensitivity, and 93.85% specificity.

4.6. RQ4: What Training Methodologies and Evaluation Metrics Have Been Used in the Rodent Models of Epilepsy, and Which of the Developed DL/ML Models Have Been Implemented?

Previous studies have used various types of rodents to analyse epilepsy. Genetic rodent models are widely employed for studying absence seizures, with temporal lobe epilepsy and Dravet syndrome being extensively investigated (see Table 1). The number of rodents used in these studies ranges from 3 to 45 for model development. The number of mice used in developing ML and DL algorithms in recent studies is shown in Table 3 and Table 4.

Various training strategies have been observed in the literature. Most studies relied on the train–test split [102,103,106,108,110,111,118]. Shin et al. [100] and Buteneer et al. [109] used ten-fold cross-validation, but Ramirez et al. [120] added an independent test set to the disjointed ten-fold cross-validation. Nandan et al. [119] and Wang et al. [101] employed two-fold cross-validation. Leave-one-out cross-validation was used by Liang et al. [107] and Baser et al. [105]. Lu et al. [104] utilised leave-one-out cross-validation with an independent test set. Singularly, Besne et al. [112] split the dataset in 100-fold cross-validation with independent test data. Some recent studies adopted train–validate and evaluate with independent test sets [34,113,114,115]. Two studies did not specify the training and validation type employed in their study [116,117].

Figure 6 shows the number of eligible articles employing different performance valuation metrics. The studies in this survey employed more than one evaluation metric. The majority of the studies included sensitivity (15), accuracy (13), and precision (12). The rest of the metrics in Figure 6 were used by fewer studies, including false-alarm rate (6), AUC (4), precision (3), and delay latency (3).

Amongst the studies included in this survey, only one study integrated the derived machine learning model into a web server for rodent models of epilepsy [34]. Wei et al. implemented an XGBoost-based model into a web server known as Epi-AI. Clinicians can upload EEG signals in EDF, CSV, and pickle formats. Epi-AI returns a list of detected seizures, the start time, and the duration of each seizure segment.

5. Discussion

5.1. Research Gaps

The increasing accessibility of computational resources, along with the limitations of manual and thresholding methods in detecting epileptiform discharges, contributes to the advancement of ML and DL techniques in this field. Most attention has been directed towards applying ML and DL to analyse human EEG seizure signals for healthcare delivery. Various categories of ML/DL techniques have been exploited. Considering recent studies, feature-based ML algorithms include random forest [124,125,126,127], gradient boosting [125], extra tree [125], SVM [124,125,128,129,130,131], naive Bayes [125], XGBoost [127], linear regression [129], and k-nearest neighbour [129]. Deep learning techniques include artificial neural network [129], long short-term memory (LSTM) [129,132], CNN [127,132,133,134,135,136,137,138], modified-CNN [138], tabNet [127], recurrent neural network (RNN) [139], and temporal convolution network [137]. More recent is the use of hybrid models, leveraging the strength of more than one algorithm to overcome individual limitations and build a robust model; these include CNN-LSTM [132,137], RNN-LSTM [133,135], CNN-SVM [138,140], simple CNN with modified-CNN [138], and LSTM with multi-scale atrous-based-DCNN [141]. Due to impending preictal and interictal class imbalance, Georgis-Yap et al. [137] developed unsupervised deep learning seizure models for predicting seizures. These models, known as anomaly detectors, are based on CNN autoencoder, CNN-LSTM autoencoder, and TCN autoencoder trained on interictal discharge for detecting preictal discharge as an anomaly. These models as applied to human EEG have demonstrated impressive performance on the test sets. However, models developed with human EEG signals obtained through noninvasive EEG, which typically have a range of channels from a single channel to 64 or 128, fail to generalise rodent EEG signals obtained from invasive EEG with at most 2 channels. Most articles on rodent models for seizure detection have used SVM compared to other algorithms applied in a single study [107,110]. However, since 2019, impressive performance has been achieved by applying complex algorithms such as XGBoost, MLP, and GNN (see Table 3). Interestingly, XGBoost is also used for seizure detection in human EEGs. Recently, DL has been investigated for seizure detection in rodent models of epilepsy, such as CNN (see Table 4). CNN methods are also commonly used for seizure detection in human EEGs. However, the utilisation of hybridised models in rodent models of epilepsy has not yet been investigated. Advancing ML and DL methods in identifying epileptiform activity in rodent models of epilepsy provides notable benefits, such as automated detection of epileptiform patterns with improved performance, early prediction of seizures, effective management of extensive datasets, and support for longitudinal studies.

A growing number of articles have not been published regarding the detection and prediction of seizures in rodent models of epilepsy, as shown in Figure 4. This may be due to data restrictions or a lack of access to the rodent EEG signal repository. On the contrary, the seizure detection and prediction algorithms developed by research during the last five years, based on human EEG data, have mostly depended on publicly accessible datasets (University of Bonn Database [142], CHB-MIT Scalp EEG [143,144], Melbourne–NeuroVista seizure trial [145], Kaggle UPenn and Mayo Clinic’s Seizure Detection Challenge, TUH EEG Seizure Corpus [146], Helsinki University Hospital EEG Database [147,148]). Although an essential precaution for patient safety is the use of rodents in research, owing to the confidential nature of the research that produces the rodent EEG data, preclinical EEG data sharing is not yet a common practice [149]. While equitable, ethical, and efficient use of shared data remains a concern, data sharing or collaboration with data scientists may provide preclinical researchers with the external expertise to explore the EEG signals and uncover patterns that might not have been originally envisioned. Preclinical investigations, combined with big data analysis of the EEG signals, can lead to swift translational research with the endpoint of producing promising new therapies, alleviating the burden of epilepsy in the general population.

A significant burden of epilepsy is the uncertainty of determining an impending seizure [150,151] in people with refractory seizures [152]. Several lines of research support the idea that seizure generation, or ictogenesis, is not random [153,154]. Classifying the interictal and preictal discharges is equal to predicting seizure symptoms according to the categorisation of the epileptiform discharges [155]. Prediction of seizures is an emerging area of clinical research which will not only assist the epileptologist in identifying the initiation of seizures but will also help provide early treatment and control of seizures. Seizure prediction has garnered attention with human EEG signal [22,152,156,157,158] to develop a closed-loop stimulator. A closed-loop stimulator aims to identify the preictal discharge and stimulate and return the brain to a stable state [159]. The clinical recordings of patients being assessed for epilepsy often last only a few days or a week. The difficulty in capturing spontaneous seizures and long-term continuous electrophysiology has greatly impeded the development of seizure forecasting [160]. Moreover, drug discovery and development is a systematic process that involves several stages before translation into healthcare. One of the stages is preclinical development. This stage aims to understand the mechanism of action of the candidate drugs, assess their potential toxicity, and verify their effectiveness using various laboratory and animal models [161]. Exploring seizure prediction with prolonged EEG recording from rodent models of epilepsy is a feasible and valuable approach to initiate and evaluate new therapeutic approaches that can control and prevent refractory seizures from occurring [162]. Furthermore, preclinical researchers will be provided real-time notification of seizure onset for optimal experimental evaluation in rodents included in the experiments. Figure 5 indicates the lack of research in seizure prediction in rodent models of epilepsy.

Rodent models of epilepsy are used to simulate human epilepsy before seizures are detected or predicted. Seizures are first classified as general, focal, or unknown onsets. The classification of seizure or epilepsy type directly affects the selection of optimal treatment [39,163]. Similarly, the development of ASDs is tailored to a particular seizure. In the recent classification of seizures and epilepsies by ILAE, published in March 2017, both generalised and focal onsets display similar types of seizures; focal onset seizures now encompass seizures (atonic, clonic, tonic, tonic–clonic, myoclonic, and epileptic spasms) that were previously thought to be associated with generalised onset. Mohammadpoory et al. [115] detected seizures in a clonic seizure dataset, but it is unclear if the seizures are of focal or generalised onset. Developing algorithmic models is the main emphasis of ML/DL seizure detection and prediction. Nevertheless, reporting the onset of seizures in conjunction with the seizure type is essential for identifying studies focusing on the seizure type of a particular seizure onset. In children and adults, focal onset seizures are more prevalent than seizures of generalised onset [164,165,166,167]. Significant correlations exist between focal epilepsy and the frequency of drug-resistant epilepsy [168]. Amidst this prevalence, most studies included in this survey focused on seizure types (absence and generalised tonic–clonic seizures) of generalised seizure onset [103,105,106,107,108,109,110,111,113]. Furthermore, epilepsy and refractory epilepsy affect people of all ages, with a high incidence and prevalence in children compared to adults/mixed age studies [168,169,170]. Although the incidence of epilepsy in children declines as they age, epilepsy in children is highest in the first year of life [167]. Thus, it negatively impacts their quality of life and the family’s functionality [171]. Absence seizures usually begin in childhood, between the ages of four and ten, even though it has been documented in individuals of all ages [172]. Detection of absence seizures has been extensively studied [103,105,106,107,108,109,110,111], in addition to few studies on Dravet syndrome. No study has aimed to detect or predict seizures in neonatal (benign familial neonatal epilepsy) and other epilepsy syndrome during infancy (West syndrome), childhood (Lennox–Gastaut syndrome, Landau–Kleffner syndrome), and adolescence (juvenile myoclonic epilepsy).

Linear features have been widely applied in detecting seizures in rodent models of epilepsy [34,102,103,106,107,108,109,110,112,115,120]. Despite the good performances obtained from using these features, linear features cannot identify the EEG changes that preceded seizures because the brain is a nonlinear dynamic system [90]. Given the nonlinear complexity of the brain, it is suitable to analyse the changes in activities of the brain with nonlinear features [173]. Nonlinear energy [111] and approximate entropy [101,107,108] have been used together with linear features. However, the approximate entropy values for the ictal and interictal signals overlap; hence, approximate entropy has to be combined with a complementary feature [107]. The dynamism of the brain signal originates from the change in statistical and frequency components over time [174], hence the dynamical property of the EEG signal [173]. Previous studies have focused on extracting features from the time and frequency domains. The time and frequency domains capture the amplitude over time and the frequency components, respectively. These two domains fail to capture the dynamicity of the EEG signal frequencies [174]. Time–frequency domain analysis has been shown to be an effective tool for studying dynamical signals since the time and frequency information is integrated into a single representation [175].

The benefits of feature engineering include enhanced model generalisation, reduced memory usage, increased computing efficiency, and improved learning performance [91]. Hence, feature selection and transformation are recognised as effective dimensionality reduction approaches. From Table 2, four studies reduced feature dimension using feature transformation techniques [102,107,118,120]. Yhis dimension reduction approach captures the underlying patterns and relationships within the dataset, improving the ML models’ performance. However, the loss of some of the original information in the feature set and the inability to interpret the linear combination of the original features [176] lead to a lack of optimal discriminative features for interpretative classification, but they are suitable for exploratory data analysis [177]. On the contrary, feature selection involves selecting a small subset of features from the original feature set, thus preserving the interpretation and leaving the original feature set unchanged [178]. Four studies in this survey [103,109,112,115] explored feature selection. The forward feature selection, a wrapper-based feature selection, uses a search algorithm to iteratively determine the optimal feature to include in the set of selected features; a modelling algorithm acts as a black box to assess the quality of the feature subset. However, with a high-dimensional feature, where d is the number of features, the search space grows by $2^d$, leading to computational intensity in finding the optimal subsets of features [91]. Furthermore, even with cross validation, the feature subsets exhibit bias towards the modelling algorithm they were assessed by, causing a dependency or overfitting on the modelling algorithm used [91,178,179]. It is thus necessary to evaluate the quality of the selected feature subset against another modelling algorithm using an independent validation sample [178]. Embedded feature selection methods integrated into or as an extended functionality of some learning algorithms to perform feature selection during model training should also be explored. Some embedded methods execute feature weighting using regularisation models with objective functions to minimise fitting errors [178]. The embedded methods are computationally efficient compared to wrapper methods; they capture the dependencies between features in the feature set and between features and the target variable [180]. Another feature selection approach, the hybrid method, emerged to address the feature selection inconsistency with several existing feature selection methods. Small changes to training data can significantly impact feature selection outcomes, particularly for small-sized, high-dimensional data [91]. Hybrid feature selection integrates the strengths of other feature selection methods. The feature subset from different feature selection methods provides a comprehensive evaluation of feature importance and increases the probability of discovering the optimal feature subset that is robust to overfitting [181]. A variety of hybrid feature selection algorithms have been proposed: hybrid of genetic algorithm and embedded regularisation [182], mutual information maximisation and an adaptive genetic algorithm [183], two-staged ant colony algorithm [184], binary state transition algorithm and ReliefF algorithm [185], and fuzzy random forest-based feature selection [186]. These hybrid feature selection algorithms and more may be investigated to discover the optimal features for discriminating epileptiform discharges in rodent models of epilepsy.

As shown in Table 3, various conventional ML algorithms have been used to identify epileptiform discharges in EEG of rodent models of epilepsy. However, SVM has garnered greater focus than other ML techniques. The ML and DL techniques have been used to model the detection of epileptiform discharges as a strict binary classification task. This task involves distinguishing between different pairs of epileptiform discharges (such as interictal vs. ictal, baseline vs. ictal, and interictal vs. preictal) from fixed-length segmented EEG signals, referred to as epochs. These epochs can either overlap or not. However, a fundamental constraint of this method is the inability of the trained binary classification model to accurately identify samples that contain both partial interictal and ictal events [187]. Typically, binary models trained for epileptiform discharge detection are labelled as seizure-event or nonseizure-event. The trained model would inaccurately classify the crossover sample as either seizure-event or nonseizure-event. Due to the uncertainty associated with crossover samples, rather than assigning EEG segments to a certain class, classifiers based on predicted probability and a threshold to classify the epileptiform discharge event may overcome the uncertainty of dealing with overlapping events. Although some of the ML/DL methods (MLP, GLM, XGBoost) employed in the literature are known probabilistic classifiers, these classifiers, when trained under an appropriate loss function, could model epileptiform discharge detection as a binary probabilistic classification task, eliminating the necessity for postprocessing.

Epileptic seizures are infrequent, second- to minute-long episodes that disrupt hours, days, or even weeks of seemingly normal brain activity in the EEG [188]. The EEG dataset is substantially imbalanced due to the wide disparity between nonseizure and seizure events [189,190]. Generally, class imbalance is a challenge in biomedical signals [191]. The seizure detection process is challenged by the intractable issue of imbalanced class distribution. This phenomenon arises due to the inherent bias of the classification algorithm towards favouring the majority class, which in this case is the nonseizure data [192]. To tackle the class imbalance issue, the studies included in this survey balanced the training data by reducing the number of nonseizure events through downsampling. Contrary to downsampling, oversampling involves making exact copies of the minority class, which can lead to overfitting [193,194]. Moreover, it is important to emphasise that oversampling does not yield any supplementary data and, hence, does not successfully address the fundamental issue of inadequate data; this explains why just increasing the sample size of the minority class does not effectively improve the identification of that class [194,195] and why downsampling the majority class seems more beneficial than oversampling [196]. However, downsampling can lead to a fabricated balance that does not precisely reflect the true distribution of categories in real-world scenarios [197]. This may lead to a model that demonstrates exceptional performance on a training set with evenly distributed data but performs poorly when presented with imbalanced real-world data. Lack of sufficient training data hinders detecting epileptiform activity in rodent models of epilepsy. Downsampling further decreases the data size for the majority class, decreasing the statistical power and dependability of the model’s classifications [198]. Additionally, when the data size is reduced, the model’s classification can become more responsive to changes in the training data, resulting in heightened variability in classification and a less reliable model [199]. Therefore, cost-sensitive learning may improve performance measures for the minority class. Cost-sensitive methods effectively handle class imbalance by imposing higher penalties for misclassifying the minority class. When identifying epileptiform activity, it is essential to prioritise signals from the ictal and preictal stages over those from the interictal and postictal states. Misclassifying ictal or preictal signals carries more severe consequences than misclassifying interictal signals. Although interictal events offer useful insights into brain activity during the intervals between seizures, the identification of ictal and preictal events is probably more essential due to their immediate and substantial impact. By prioritizing the identification of ictal or preictal events, model performance may improve amidst class imbalance, providing preclinical researchers with the capacity to investigate the underlying mechanisms of epilepsy, predict seizure onset, and evaluate the impact of drug trials on rodents.

In this review, SVM, ANN, and DL models in rodent epileptiform discharge detection have shown outstanding performance, as seen in Table 4. The prevailing notion is that complex ML/DL models offer enhanced performance [200] as a result of their complex computational capability in discovering nonlinear relations in EEG signals and several application domains. These models are frequently considered black boxes [201], which are nontransparent and prove difficult for nontechnical people to understand their decision-making process. Government agencies have debated the lack of transparency regarding the design, development, evaluation, and implementation of AI tools. This might not be a challenge in some fields, but the performance of algorithms and the ease of understanding ML and DL models are very important for their use in the medical domain [202], which results in model transparency that enhances the trust of the users. However, there is often a compromise between how well a model works and how easy it is to understand [203]. High-performance models are often more complex and opaque, while interpretable models may not have good performance [204]. As part of the complete data science procedure, performance and interpretability metrics should be iteratively refined to update other steps of the process [205]. Consequently, Rudin et al. [206] suggested interpretability as beneficial for troubleshooting in practical scenarios, leading to improved rather than diminished performance. Furthermore, interpretability should be prioritised for high-stakes decisions like criminal justice, medicine, and banking rather than using post hoc models to explain black-box credibility [206,207]. Although interpretability and explainability are often used interchangeably, post hoc explainability techniques aim to audit black-box ML/DL models by creating simplified surrogate models [208]. By extension, nontechnical people can use these simplified surrogate models to understand the inner workings of black-box ML/DL models [209,210]. However, concerns have been raised regarding the effectiveness of post hoc explainability methods in elucidating the inner mechanism of black-box models, and certain drawbacks have been observed [200,207,208,211,212,213]. Amid these concerns, numerous ML/DL approaches have been developed that maintain high accuracy while achieving significant interpretability, even on the most challenging datasets [214,215,216,217,218,219]. Hence, the interpretable detection of epileptiform discharge in rodent models of epilepsy should be prioritised. It is also important to involve preclinical researchers as stakeholders in the development of these models rather than developing them solely within the field of computer science. This ensures that their requirements are taken into consideration.

Regardless of the training methodology adopted in training ML/DL models, data splitting should be disjointed such that all epochs of a mouse belong to the same set. Only one study clearly stated how the signal epochs were split without intersection [120]. In addition to training data, training methodology or data splitting is critical to the performance of the ML/DL models. As is well-known, the performance of a model on the training set is an unduly optimistic prediction of its performance on unseen data. It is, therefore, a common practice to set aside some data to evaluate the performance of a model. The most straightforward and probably the most frequently used training strategy has been the train–test split [220]. The seven studies in this survey that adopted the train–test split methodology ensured the epileptiform discharge classes were balanced [102,103,106,108,110,111,118] despite the inability of this data-splitting methodology to maintain a balanced representation in the train sets and EEG data being highly unbalanced, resulting from the scarcity of epileptiform discharges in EEG signals. Furthermore, the train–test split is less efficient with limited data. With the limited number of mice in the studies, using train–test split is unsuitable and risks overfitting/underfitting. Additionally, different splits may lead to varying model performances [221]. On the contrary, K-fold cross validation reduces the dependence of the model on the single source of training data, and a stable model estimate is achieved by averaging the performance of the different folds. K-fold cross validation was adopted in eight studies in this survey; the most frequently used is ten-fold cross-validation. This training methodology has no defined rule for determining the optimal value of K. The degree of variance and bias brought about by the training process and the selection of K are subject to trade-offs. Hence, a range of K values should be tested by treating the value of K as a hyperparameter tuned to optimize the performance of the model.

The models in this survey cannot be directly compared owing to the nonuniformity of the evaluation metrics reported in the respective studies. The main metrics of interest in epileptiform discharge detection are sensitivity (the ability to detect seizures), false-alarm rates (false-positive rate) [222], and delay latency (delay between EEG onset marked by experts and detected seizure by automatic detection system) [187]. As shown in Figure 6, the majority of the studies in this survey evaluated the performance of their model with high sensitivity, accuracy, and specificity. Only a few reported the false-alarm rate. Unfortunately, in epileptiform discharge detection, a high specificity does not always equate to a low false-alarm rate [187]. Moreover, zero or shortest detection latency is essential for accurate seizure detection and prompt clinical intervention. However, most studies ignored these critical metrics. Notably, reporting delay latency will further promote research on developing methods to reduce delay latency, leading to optimal epileptiform discharge detection models.

While most of the models included in this review remained experimental, only one study deployed the ML model as a web server. Experimental ML/DL models are important for exploring new techniques and advancing epileptiform discharge detection. These models are often fine-tuned to achieve remarkable performance on the training data but have difficulties in generalising to real-world data, which may possess noise, incompleteness, and diversity. Without validation on real-world datasets, the efficacy of these models, when used in real-world applications, may be compromised. The lack of deployment for real-world validation limits the practical applicability, reliability, and scalability of the experimental models. It is essential to ensure that these models can function reliably in real-world applications rather than solely in a controlled laboratory environment.

5.2. Limitations of the Study

The search conducted for this evaluation included the time period from 1 January 1994 to 1 January 2024. This study does not include papers published later or beyond the specified date range. Our review may lack current and potentially crucial findings if there is recently published research. Furthermore, studies that were not performed in English were excluded. This could result in bias, as important discoveries from countries with English as a foreign language may not be represented. Furthermore, journals that were not located in the JCR and conferences that were not indexed by ACM or IEEE were omitted. Consequently, articles with noteworthy findings could be found in other journals and indexes. Although we performed a comprehensive and methodical search in the two databases (PubMed and Google Scholar), there is a possibility that we may have overlooked relevant articles due to limitations in search terms, database coverage, or indexing problems.

5.3. Future Works

Future reviews may overcome these constraints by integrating a wider array of sources, especially publications in languages other than English, and implementing more uniform techniques throughout investigations. Although there are limitations, the results of this review offer insights and contribute to the continuing discussion on the automatic detection of epileptiform discharges on the EEG of rodent models of epilepsy. Furthermore, an essential area of study is evaluating the efficacy of machine learning and deep learning in silico methods for accurately predicting the consequences of a genetic variant in epilepsy.

6. Conclusions

The use of rodent models in epilepsy research provides invaluable insight into the mechanisms underlying the development of epilepsy in human brains. Additionally, these models have been instrumental in conducting preclinical studies to determine the effectiveness of potential therapeutic interventions. Utilising ML/DL models for epileptiform discharge in rodent models of epilepsy significantly reduces the time and effort required for the manual analysis of EEG signals. Therefore, researchers can focus on other essential aspects of epilepsy research. This systematic review examined and summarised the last three decades of research on detecting epileptiform discharge in rodent models of epilepsy using ML/DL techniques. This review is intended for researchers in the field of EEG epilepsy detection in rodent models of epilepsy who seek to address the existing gaps in the literature about the approaches proposed in recent years. To the best of our knowledge, there has not been a literature review regarding the ML/DL techniques for detecting epileptiform discharge in rodent models of epilepsy. This study focused on seizure type, features and feature engineering, ML/DL methods, training methodologies, evaluation metrics, and model deployment. We discovered that seizure detection receives more attention than seizure prediction. Furthermore, the number of mice included in the studies was limited, invariably leading to insufficient training data. Amidst the prevalence of drug resistance in focal epilepsy, absence seizures and generalised tonic–clonic seizures have received more attention. Although time–frequency and linear features are proven effective tools for studying dynamic signals, this crucial feature domain has not been utilised. Probabilistic classification has not been appropriately explored to model crossover EEG signal samples. With the high performance of the models, delay latency, which is a crucial metric in evaluating epileptiform discharge detection, has been greatly ignored. Except for one model, all other models are experimental and lack real-world validation. As important as interpretability is in the biomedical domain, there is a scarcity of intrinsically interpretable epileptiform discharge detectors. Given these limitations in the current studies, further studies are required to develop, deploy, and validate ML/DL models for rodent epileptiform discharge in real-world settings.