Our systematic review was implemented following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [17] in Multimedia Appendix 1 and prospectively registered with PROSPERO (ID CRD42023467260).

We developed detailed inclusion and exclusion criteria for our systematic review based on population, modeling, study type, language, and outcome measures, as detailed in Textbox 1.

PubMed, Cochrane, Embase, and Web of Science databases were thoroughly retrieved as of August 27, 2023. Both MeSH (Medical Subject Headings) terms and free-text keywords were used, without restrictions on publication region or year. Details of the search strategy are available in Table S1 in Multimedia Appendix 2.

All identified articles were imported into EndNote software (Clarivate). After deleting duplicates, titles and abstracts were screened to rule out irrelevant studies. The remaining articles were reviewed in full for inclusion. A spreadsheet was used to extract data, including title, first author, year of publication, type of study, patient origin, dataset source, number of epileptic seizure cases, total number of epileptic seizure cases in the training set, overfitting methods, number of cases in the validation set, missing value handling methods, variable selection, types of models used, and modeling variables.

In total, 2 independent investigators (ZZ and BC) implemented the literature screening and data extraction, followed by cross-checking. In case of any disagreements, a third investigator (DX) was consulted for resolution.

The QUADAS-2 tool (Quality Assessment of Diagnostic Accuracy Studies–2) [18] was leveraged to appraise the risk of bias and applicability of the included studies. This tool evaluates 4 aspects, such as patient selection, index test, reference standard, and flow and timing. Each domain contains specific questions answered as “yes,” “no,” or “uncertain,” corresponding to a bias risk of “low,” “high,” or “uncertain,” respectively. Studies were considered at low risk of bias if all key questions in each domain were answered with “yes.” Any “no” response indicated potential bias, requiring evaluators to assess the risk level according to established guidelines. An “uncertain” rating signified insufficient information for a definitive judgment.

This systematic review assessed diagnostic accuracy through sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, diagnostic odds ratio, and the summary receiver operating characteristic curve. Calculating these estimates required diagnostic 4-fold tables. For ML, given the risk of overfitting, the diagnostic 4-fold tables in both training and validation sets were extracted from studies on ML. Funnel plots were used to analyze the publication bias of the C-index and then Egger test was performed for statistical evaluation of the publication bias. For DL, only the results of validation or test sets were considered.

Data analysis was executed using Stata 15.0 (StataCorp LLC). A bivariate mixed-effects model was used for the meta-analysis of sensitivity and specificity. However, if original studies lacked diagnostic 4-fold tables, they were calculated using two methods: (1) one based on sensitivity, specificity, precision, and case numbers, and (2) the other based on sensitivity and specificity calculated using the optimal Youden index and case numbers. Pooled estimates of sensitivity, specificity, likelihood ratios, diagnostic odds ratio, and their 95% CIs were calculated. The summary receiver operating characteristic curve area was also estimated. Publication bias assessment was conducted using funnel plots. Statistical significance was determined at P<.05. The studies included were all based on EEG-based ML methods, and we know that traditional ML and DL have certain differences in their image-processing capabilities and modeling efficiency. Therefore, to reduce heterogeneity, we conducted a subgroup analysis by model type for traditional ML and DL.

All analyses were based on previously published studies; thus, no ethical approval and patient consent are required.

Our systematic search identified 15,389 related articles, of which 3130 duplicates (2975 identified by software and 155 manually) were excluded. After title and abstract screening, we excluded 12,014 additional articles. Of the remaining 44 articles, 6 unpublished conference abstracts and 2 articles without full-text access were excluded. Among the remaining 36 articles, we excluded 5 for missing outcome measures, 1 for data overlap, and 2 for including adult participants. After these exclusions, 28 studies met the eligibility criteria [19-46]. Figure 1 illustrates the literature screening process.

Table 1 presents the detailed characteristics of the 28 included studies, with 15 on ML [32-46] and 13 on DL [19-31]. Of these, 6 were multicenter studies [23,26,27,32,34,39], while the remaining 22 were single-center investigations. Furthermore, 5 studies were prospective [37,40,41,43,44], with the remaining 23 being retrospective. A total of 6 studies incorporated prospective external validations [26,27,30,40,43,44], 2 implemented multicenter external validations [26,27], and the remaining 22 used internal validations. Among studies with internal validation, 3 used 10-fold cross-validation [25,32,34], 4 used 5-fold cross-validation [21,35,37,41], 3 used the leave-one-out method [22,32,45], 1 used the Bootstrap method [33], and 10 used random sampling [19,20,23,28,31,36,38,39,42,46]. Of the total, 1 study did not specify the generation method of validation set [24]. Across the 28 studies, 5742 participants were included, with 1814 confirmed patients with epilepsy, ranging from 9 to 1117 patients per individual study. Furthermore, 10 studies were carried out in Asian countries, while 18 were conducted in Western countries. Out of total, 8 studies used EEG signals from the CHB-MIT dataset [19,22,24,31,34-36,39], another 8 used signals from The University of Helsinki dataset [20,21,23,25,28,29,38,42], and 3 used data from Cork University Maternity Hospital [26,27,45]. In addition, 1 ML study used 3 distinct models, each with separate training and validation sets [38]. All participants in the included studies were children, with 19 studies focusing specifically on neonates [19-21,23,25-30,32,33,36,38,42-46]. Studies on DL predominantly used CNN models. Of the 13 studies on DL detection of pediatric epileptic seizures included in this paper, 7 studies mainly applied the CNN model [21-23,25,27,29,30]. The DL models also included adaptive grey wolf optimizer (AGWO) [19], graph convolutional neural network [20], convolutional gated recurrent neural network [31], cross-feature fusion stream convolutional neural network [24], and pretrained deep convolution neural networks [28], and these models are mostly evolved from the CNN. ML studies used models including random forest (RF), k-nearest neighbors (KNN), support vector machine (SVM), extreme gradient boosting, adaptive boosting, and decision tree (DT). RF is an ensemble learning algorithm used for classification and regression [47], and 1 of its primary advantages is its robustness to noise. In our study, 4 original studies using this model were included [32,33,37,39]. KNN is the simplest ML tool with supervision for classification [48], and it depends on the predefined nearest number or the K value. One of the included original studies used this model [36]. DT predicts the correct classification by recursively partitioning the real space, and it consists of 2 types of multiple nodes, that is, leaf nodes and decision nodes. Of the total, 2 included original studies used this model [33,36]. SVM is an advanced algorithm mainly used for pattern recognition and feature reduction, using discriminative techniques to classify the input, and is best suited for binary classification [49]. Up to 5 original studies on ML in this study applied this model [34,38,42,44,45]. With the exception of 1 study that used electrocardiography as a variable [35], all other studies used EEG data. Clinical characteristics of patients were included as modeling variables in 6 studies [32,37,40,41,43,46].

In the 28 included studies, all cases were consecutive case series, eliminating concerns of case selection bias. Notably, the type of study design had minimal influence on the results of image-based DL studies [21,26,27], leading to a low risk of bias for this category. Conversely, conventional ML methods often involve manual segmenting regions of interest segmentation and extracting texture features, potentially introducing significant bias in case-control studies. Consequently, this study was categorized as having an unclear risk of bias [35]. There was a study [32] in which we doubted its results were interpreted without the outcome of the gold standard trial of epileptic seizures, the risk of bias of the study was classified as unclear. Further details are shown in Figure 2.

Real-time detection of epileptic seizures has profound clinical significance. Studies have shown that if a child, who has a normal nervous system without a history of neurological diseases, experienced unprovoked epileptic seizures without obvious acute causes, then the risk of recurrence is about 25% within the following 1 year, and 45%-50% within the following 3 years [50-52]. In this study, the results showed that AI-based methods have high sensitivity, specificity, and accuracy in detecting epileptic seizures. The overall C-index, sensitivity, and specificity of ML are 0.73 (95% CI 0.67-0.79), 0.88 (95% CI 0.83-0.91), and 0.83 (95% CI 0.71-0.90) in the validation set, respectively. DL exhibited even higher accuracy, with an overall C-index, sensitivity, and specificity of 0.91 (95% CI 0.88-0.94), 0.89 (95% CI 0.85-0.91), and 0.91 (95% CI 0.88-0.93) in the validation set. It was also found that EEG-based DL appears capable of detecting SE in a timely manner. This suggests that this study has the potential to be used in the development of portable devices for real-time monitoring in future research. Such devices could enhance the timely detection of epileptic seizures, enabling prompt clinical intervention and thereby reducing the risk of recurrence.

Several noninvasive methods for epilepsy detection are actively investigated. In recent years, researchers have also explored other noninvasive means for detecting epileptic seizures besides EEG, such as magnetoencephalography (MEG). MEG measures the magnetic fields generated by the brain to provide information about neural activity during epileptic seizures. Similar to EEG, MEG can also be used to detect signal changes before epileptic seizures. Some studies have shown that MEG has specific advantages in detecting extratemporal epileptic spikes, especially those from epileptic lesions located on the surface of the brain [53-55]. A meta-analysis by Brændholt and Jensen [56] showed that the sensitivity and specificity of MEG for detecting epileptic seizures were 0.77 (95% CI 0.60-0.90) and 0.75 (95% CI 0.53-0.90), respectively. While this finding supports MEG as an accurate method for detecting SE, compared with the results of this study, it suggested that both EEG-based ML and DL may offer superior performance in detecting SE. Another method is electrical source imaging (ESI), which is a model-based imaging technique that integrates the spatial and temporal components of EEG to identify the sources of abnormal electrical activity related to epileptic seizures [57]. In a study by Ricci et al [58], the application of ESI in pediatric patients with refractory epilepsy was evaluated. The results showed that the sensitivity, specificity, and accuracy of ESI in predicting seizures in children were 0.57 (95% CI 0.34-0.78), 0.86 (95% CI 0.57-0.98), and 0.69 (95% CI 0.51-0.84), respectively (P=.01). Furthermore, 1 study showed that the diagnostic performance of ESI is comparable with that of magnetic resonance imaging and positron emission computed tomography. The sensitivity of these 3 methods for predicting epileptic seizures was 0.88, 0.71, and 0.66, respectively, while their specificity was 0.47, 0.71, and 0.59, respectively [59]. While these detection methods have shown promise, real-time monitoring of epileptic seizures remains a challenge, requiring further attention in clinical practice. Our study demonstrates that DL-based approaches offer promising results with a sensitivity of 0.89 (95% CI 0.85-0.91) and a specificity of 0.91 (95% CI 0.88-0.93) for pediatric epileptic seizure detection. Integrating DL based on EEG or electrocardiography into portable smart devices could significantly enhance real-time monitoring capabilities.

We have also noticed that there are permanent differences between DL and ML, and both have their own advantages. For ML, interpretable clinical features and corresponding ML models can be developed. For example, one of the original studies we included based on ML showed that the RF, SVM, and KNN models had a sensitivity of 0.93, 0.76, and 0.90, a specificity of 0.89, 0.76, and 0.90, and an accuracy of 0.91, 0.81, and 0.91 in predicting epileptic seizures, respectively [39]. Interpretable models in the clinical setting are still a direction that we are very interested in. However, for certain imaging applications, ML requires manual preprocessing, like segmenting regions of interest and extracting features. During this process, it is difficult to avoid any heterogeneity or bias due to manual experience, therefore, for image analysis, we prefer to use the DL to intelligently train the differences between them to develop intelligent detection tools. The primary advantage of CNN is that it can automatically detect important robust features without any manual intervention. The point is to estimate and determine the number of layers and the size of the filters in each layer, and the depth of the structure is the key to deep CNN. The network depth change will affect the convolutional receptive field and the corresponding learning feature complexity [60]. For instance, one of the original studies based on DL showed that the model had a sensitivity of 0.96, a specificity of 0.97, and an accuracy of 0.97 in predicting epileptic seizures, and was compared with common ML methods, such as RF, DT, and SVM [19].

In clinical practice, we need real-time monitoring for some diseases, such as epileptic seizures. However, ML still faces serious challenges in real-time disease monitoring, as the ML modeling is based on manual encoding, making it difficult to process images intelligently. Compared with ML, DL has significant advantages in real-time monitoring. We also found that researchers have done further work in other fields, developing wearable devices for real-time disease monitoring based on DL methods, such as monitoring heart-related diseases [61,62] or automatically assessing the severity of knee osteoarthritis [63].

We found that current meta-analyses related to epilepsy detection are usually based on the direct diagnosis by clinical neurologists through the reading of EEG signals [64,65]. This study is the first study to discuss the use of AI-based methods, including ML and DL, for the prediction and diagnosis of epileptic seizures. Particularly, DL has shown very desirable accuracy in the diagnosis of SE, which may provide some theoretical support for the subsequent development of intelligent reading tools or wearable devices. However, several limitations should be considered. First, the number of cases used for model construction is limited, whereas ML algorithms require a larger sample size to build robust models. This may influence the generalizability of these results. Second, some included studies lack independent validation. Third, the detection performance of DL across different age groups remains uncertain. Third, we only searched for literature published in English in this study. For this, there are 2 primary reasons, one is there are certain search barriers in this study and the other is that considering most readers would need to review the relevant original studies while reading this article, we developed this search strategy to improve the readability for the readers. Of course, it is also a limitation of this study, and we hope that future research can cover more detection tools developed by diverse populations from different countries and ethnicities to validate the results of this study.

Our systematic review demonstrates promising accuracy of AI methods in epilepsy detection. DL appears to offer higher detection performance compared with ML, this finding supports our initiative to further research and develop early warning tools using DL.