Predicting response to neuromodulation therapies in drug-resistant epilepsy using machine learning models: a meta-analysis and systematic review

Abstract

Objective

The aim of this study was to identify and analyze all the relevant literature regarding the use of machine learning to predict response to neuromodulation therapies in patients diagnosed with drug-resistant epilepsy.

Material and methods

We systematically search PubMed, Embase, Scopus and Cochrane databases to identify all the studies that used machine learning models to predict response to neuromodulations. Prior to the search, the study was registered at the International prospective register of systematic reviews (PROSPERO, CRD42024543952). Quality assessment and risk of bias was done using PROBAST. A random effects model was used to calculate the pooled value of the AUROC. A sub-analysis was performed for population-specific scenarios.

Results

A total of 4,451 studies were identified after our initial search, from those, only 12 papers were included in the final analysis. The total number of patients across all the cohorts was 535. 11 studies focused on VNS and only one on ctDCS. Only five out of the 12 studies included an external cohort to validate the results. The most common population was pediatric (n = 7). The most common ML model used was the support vector machine. The pooled area under the receiver operating characteristic curve (AUROC) was 0.84 (95% IC, 079–0.88).

Conclusions

Our study suggests that multimodal ML approaches show promising performance in predicting response to neuromodulation strategies in patients with drug-resistant epilepsy. However, the limited number of studies, the scarcity of external validation and small cohorts highlight the need for larger, high-quality prospective investigations to confirm these findings and improve the generalizability of ML-based prediction models.

Supplementary Information

The online version contains supplementary material available at 10.1186/s42234-025-00191-8.

Background

Epilepsy, as defined by the International League Against Epilepsy (ILAE) in 2014, is a disorder of the brain characterized by an enduring predisposition to epileptic seizures (Fisher et al. 2014). It is estimated that around 1.2% of the United States population has active epilepsy, which accounts to approximately 3,500,000 cases (3,000,000 adults and 500,000 children) for 2021 (Zack and Kobau 2017; Epilepsy Facts and Stats | Epilepsy | CDC n.d.), with annual costs as high as 30,000USD per person, annually (Begley and Durgin 2015). Of relevance, these figures may underestimate the current burden of epilepsy. The estimate of 3.5 million cases reported by the CDC was derived using the 2021 U.S. population of about 331 million, whereas the 2024 census reports a population of 342,787,750 (Kobau et al. 2024). Of all the patients diagnosed with epilepsy, around 15–30% will not respond to at least two trials of antiseizure medication (ASM), a scenario known as drug-resistant epilepsy (DRE) (Kwan et al. 2010). In a 2021 meta-analysis by Sultana et al., the prevalence of DRE was estimated at 13.7% (95% CI: 9.2–19.0) in population- or community-based samples, but was substantially higher 36.3% (95% CI: 30.4–42.4) in clinic-based cohorts (Sultana et al. 2021). This is significant, as DRE is associated with increased costs (Ngan Kee et al. 2023; Sheikh et al. 2020), disability (Sarkis et al. 2018) and mortality (Rheims et al. 2022), compared to non-DRE patients.

While the mainstay of treatment for these patients is surgical removal of the epileptogenic focus, not every patient is candidate for this type of intervention (Jehi et al. 2022). In those cases, a palliative approach is warranted, which could include different neuromodulation techniques, in an attempt to reduce the number of seizures and improve the patient outcomes and quality of life (Englot et al. 2017). Of the available options, vagus nerve stimulation (VNS) is the most common, however, other types of neuromodulation techniques include deep brain stimulation (DBS), responsive neurostimulation (RNS), either cortical (RCS) or subcortical (RScS), and cathodal transcranial direct current stimulation (ctDCS) (Ghosh et al. 2023; San-Juan, 2021).

These techniques involve the use of different devices that deliver electrical impulses directly to the central nervous system (CNS) or to the vagus nerve (VN), to modify neuronal and field activity in order to reduce the frequency and seizure severity (Boon et al. 2018). Of relevance, due to the fact that less than 5% of the patients achieve actual seizure freedom, response to neuromodulation treatment is usually defined as a reduction in 50% in the number of seizures (Ryvlin et al. 2021).

In recent years, machine learning (ML) and deep learning (DL) have shown potential to revolutionize various aspects of healthcare (Habehh and Gohel 2021). By leveraging multimodal datasets, these models have been used to predict seizure events (Rasheed et al. 2021), determine the probability of a given patient to response to ASM (Kaushik et al. 2024; Wu et al. 2022), and enhance the efficacy and safety of neuromodulation interventions by creating closed-loop circuits (Ramgopal et al. 2014; Dümpelmann 2019).

This review aims to identify all previous studies focused on training and validating ML models to predict response to neuromodulation techniques in patients with DRE.

Materials and methods

Search strategy and registration

The search strategy was designed and conducted by an experienced medical librarian (J.M.) with input from study investigators. PubMed, Embase, Scopus, and Cochrane Library were searched from inception through July 15, 2024. Search strategies were formulated using controlled vocabularies and keyword capabilities in each database. A complementary manual search was performed on additional databases (Google Scholar) on August 15, 2024, and by checking the references of all main review articles on this topic to identify possible additional eligible studies. The detailed strategies are reported in Supplementary Table 1. Before starting, the study protocol was registered and published in the online International Prospective Register of Systematic Reviews (PROSPERO) database (ID: CRD42024543952).

Eligibility criteria

We only included studies focusing on human research. To be eligible, patients should have been diagnosed with DRE following the ILAE guidelines (Kwan et al. 2010). We included both pediatric and adults patients treated with any type of neuromodulation technique. We only included peer-reviewed English original publications, encompassing clinical trials, cohorts, case–control studies, systematic reviews and meta-analysis. Other types of publications like narratives reviews, case series, case reports, editorials, conference abstract, non-peer-reviewed and grey literature were excluded.

Regarding the models of interest, we focused on supervised classification ML and DL models that looked at response prediction. No restrictions were placed on the type or number of features included. However, all models used a common outcome label, since response to neuromodulation therapies is typically defined as a ≥ 50% reduction in seizures. Unsupervised or reinforcement learning models were excluded. For more details on the selection criteria see Table 1.

Table 1.

Selection criteria

Inclusion criteria

Population: human population, pediatric and adults patients with diagnosis of DRE by the ILAE 2014 criteria

Intervention: VNS, DBS, RNS, ctDCS, TMS, TNS

Studies: peer-reviewed studies, clinical trials, prospective and retrospective cohorts, case–control studies, systematic reviews and meta-analysis

Models: supervised classification ML or DL models

Exclusion criteria

Population: non-human population, diagnosis of DRE by other criteria

Intervention: neuromodulations techniques used alongside surgery

Studies: non-peer reviewed publications, articles in languages other than English, case series, case reports, narrative reviews, editorials and expert opinion; and grey literature

Models: unsupervised machine learning models, regression architectures

ctDCS Cathodal Transcranial Direct Current Stimulation, DBS Deep-brain stimulation, DL Deep learning, DRE Drug resistant epilepsy, ILAE International league against epilepsy, ML Machine learning, TMS Transcranial magnetic stimulation, TNS Trigeminal nerve stimulation, VNS Vagus nerve stimulation

Data extraction

The selection process involved two phases, an initial screening based on titles and abstracts, followed by a full-text review to confirm eligibility based on the predefined inclusion and exclusion criteria. Two investigators (AQ and FF) independently evaluated the studies for eligibility and resolved any disagreements through discussion and consensus. When no agreement could be reached, a third investigator (TZ) determined the eligibility.

For eligible studies, a standardized data extraction process using Convidence software was utilized. For each study, the primary outcome was the area under the receiver operating characteristic curve (AUROC). As secondary outcomes, we also extracted the following performance metrics of each model: sensitivity (recall), specificity, precision, accuracy, and f1 score. In the case of studies using the same cohort but with different features, we considered each of them as individual models. In articles that did not report these metrics but reported confusion matrices, the metrics were calculated based on them. Furthermore, in case of not reporting the standard errors (se) of the AUROC, we calculated using previously published formulas (Olender et al. 2023).

Quality and risk of bias assessment

The extracted studies were analyzed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). This tool assesses the risk of bias in four domains: participants, predictors, outcome and analysis. PROBAST also determines the applicability of prediction model studies based on three domains, participants, predictors and outcomes.

Each domain was rated for risk of bias or applicability concern in “low,” “high,” or “unclear.” Two reviewers independently applied the PROBAST tool to all included studies, with discrepancies resolved through discussion or adjudication by a third reviewer.

Statistical analysis

Data synthesis and meta-analysis was conducted using the metafor package version 4.6–0 (Viechtbauer and metafor 2025) in R version 4.4.2. Heterogeneity among studies was evaluated using the I² and Cochran’s Q test. A random-effects model was used to account for potential heterogeneity across studies. Forest plots were generated to visually display the weighted AUROC values with a 95% interval confidence (IC), as well as the pooled AUROC of all the models. Population-specific analysis was also performed as a possible source of heterogeneity.

Results

Study identification

Our search identified 4,451 studies. After removing duplicates, a total of 3,362 studies underwent title and abstract screening. From the title and abstract screen, 2,838 studies were considered irrelevant and 524 progressed to full-text review. From the full-text review, 512 studies were excluded and only 12 studies progressed to data extraction. The reasons for exclusion were wrong patient population (n = 40), wrong intervention (n = 42), wrong study design (n = 212), wrong setting (n = 210), non-peer-reviewed (n = 1) and publications in another language other than English (n = 7). Figure 1 shows the study selection flowchart based on the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 flow diagram.

Fig. 1.

PRISMA selection flowchart

Studies characteristics and modalities

With the 12 extracted studies, the total number of patients was 535, including both the initial training and testing cohort (referred to as train-test), as well as the external validation cohort (referred to as external), when available. All studies focused on VNS with the exception of one that analyzed response to ctDCS (Hao et al. 2021). Some of the studies had the same model trained on different features which we considered separately for the analysis, and thus, the total number of models was 17, with five studies (a total of seven models) including an external cohort for validation. Seven of the studies included a pediatric-only population, four adult-only and one included both pediatric and adult populations. None of the studies included DL models.

Regarding study modalities, most studies incorporated at least one modality. Electroencephalography (EEG) was the most frequently used (n = 5), followed by diffusion tensor imaging (DTI), functional magnetic resonance imaging (fMRI), and magnetoencephalography (MEG) (each n = 2). The most common type of cross-validation (CV) was fivefold (n = 8), followed by tenfold (n = 5) and finally leave-one-out (LOO) CV (n = 4). The smallest train-test cohort included 17 patients, while the biggest included 70. The majority of the trained models were support vector machines (SVM; n = 10), and the second most common being the random forest (RF) classifier with four models from the same cohort but trained with different features. Other types of models included extreme-gradient boosting machine (XGB), linear discriminant analysis (LDA) and naïve Bayes classifier (NBC; Table 2).

Table 2.

Summary

Author	Population	Intervention	Data type	Features	Set	n	CV	Model
Ibrahim et al. (2017)	Pediatric	VNS	fMRI	Correlation coefficients of connectivity	Train-test	21	5	SVM
					External	8	-	SVM
Babajani-Feremi et al. (2018)	Pediatric	VNS	MEG	Transitivity and modularity from PLV values of theta, alpha, and beta bands	Train-test	23	10	NBC
Brázdil et al. (2019)	Adult	VNS	EEG	Theta, alpha, beta, and gamma band power	Train-test 1	60	LOO	SVM
					Train-test 2	-	LOO	LDA
					`External 1	22	-	SVM
					External 2	-	-	LDA
Mithani et al. (2019)	Pediatric	VNS	EHR, DTI, MEG	wPLI of broad band (2–20 Hz), low gamma 1 (30–40 Hz), and low gamma 2 (40–50 Hz), mean FA and clinical characteristics of the seizures	Train-test	38	10	SVM
Mithani et al. (2020)	Pediatric	VNS	MEG	Spatial deviation, amplitude and latency of SEFs; and wPLI	Train-test	36	10	SVM
					External	12	-	SVM
Fang et al. (2021)	Mixed	VNS	ECG	Slope, HF LF, and VLF HRV metrics	Train-test 1	59	5	RF
					Train-test 2	-	5	RF
					Train-test 3	-	5	SVM
					Train-test 4	-	5	SVM
Hao et al. (2021)	Adult	ctDCS	fMRI	Clustering coefficient, mean shortest path length, and local efficiency	Train-test	20	5	SVM
Ma et al. (2022)	Pediatric	VNS	EHR, EEG	Clinical characteristics and PLV, PLI, and wPLI calculated in six bandwidths (delta, theta, alpha, beta, low beta and high beta)	Train-test	70	10	SVM
					External	18	-	SVM
Chen et al. (2023)	Pediatric	VNS	EEG	Edges calculated from partial correlation connectivity analysis	Train-test 1	16	5	SVM
					Train-test 2	-	5	XGB
					External 1	7	-	SVM
					External 2	-	-	XGB
Berger et al. (2023)	Adult	VNS	EEG	Slope, charge output and saturation of laryngeal motor evoked potentials	Train-test	42	LOO	SVM
Cheng et al. (2024)	Pediatric	VNS	EHR, EEG	Clinical characteristics, alpha band PLI, and entropy values	Train-test	65	10	SVM
Berger et al. (2024)	Adult	VNS	DTI	Mean and radial diffusivity	Train-test	18	LOO	SVM
Overall						535

ctDCS CV Cross-validation, DTI Diffusion Tensor Imaging, ECG Electrocardiogram, EEG Electroencephalogram, EHR Electronic health record, fMRI functional magnetic resonance imaging, HF High frequency, HRV Heart rate variability, LF Low frequency, LOO Leave-one-out, MEG Magnetoencephalography, NBC Naïve bayes classifier, LDA Linear discriminant analysis, PLI Phase lag index, PLV Phase lag value, RF Random forest, SD Standard deviation, SEF Somatosensory evoked field, SVM Support vector machine, VLF Very low frequency, VNS Vagus nerve stimulation, wPLI weighted phase lag index

Risk of bias and applicability

Only half of the studies (n = 6) were considered with low risk of bias, with the Analysis domain being the most prevalent one with high risk. The applicability was low risk in half the studies, with the Outcome domain having the greatest number of unclear categories and the Participants the one with the highest risk (Table 3).

Table 3.

Risk of bias and applicability

Green indicates low risk of bias (RoB) and good applicability; yellow with a question mark indicates unclear RoB and/or applicability; red indicates high RoB and/or limited applicability

Model performance

The pooled AUC of all the studies, including both train-test and external cohorts, was 0.84 (95% CI, 0.79–0.88), with moderate heterogeneity (I² = 43.8%; Fig. 2). The highest performances in the train-test cohorts were achieved using SVM models in pediatric populations by Mithani et al. 2019 (AUROC of 0.93, 95% CI 0.87–0.99) and Mithani et al. 2020 (AUROC = 0.93, 95% CI 0.85–1.01), whereas the highest performance in the external cohort was achieved by Chen et al. 2023 with both SVM (AUROC = 0.91, 95% CI 0.67–1.15) and XGB (AUROC = 0.90, 95% CI 0.77–1.03) models, however the standard error (se) of the AUROC was high in the SVM (se = 0.12). When focusing on specific populations we can observe a reduction in heterogeneity (I² = 20.2% in pediatrics and 18.2% in adults), with a similar AUROC for the pediatric population (0.88, CI 0.84–0.92) and lower for the adult population (0.76, CI 0.70–0.83) when comparing with overall pooled AUROC (Fig. 3 and 4).

Fig. 2.

Forest Plot of the AUC for all the models and cohorts. Red dashed line marks 0.5 AUROC (random chance of classification). NBC: naïve bayes classifier, LDA: linear discriminant analysis, SVM: support vector machine

Fig. 3.

Forest Plot of pediatric-only studies. Red dashed line marks 0.5 AUROC (random chance of classification). NBC: naïve bayes classifier, LDA: linear discriminant analysis, SVM: support vector machine, XGB: external gradient boosting

Fig. 4.

Forest Plot of adult-only studies. Red dashed line marks 0.5 AUROC (random chance of classification). LDA: linear discriminant analysis, SVM: support vector machine. Forest Plot of adult-only studies. Red dashed line marks 0.5 AUROC (random chance of classification). LDA: linear discriminant analysis, SVM: support vector machine

The overall accuracy across train-test and external models was 0.76 ± 0.11, the recall of 0.77 ± 0.19, precision 0.83 ± 0.08, specificity 0.76 ± 0.22 and F1 score of 0.83 ± 0.10 (Table 4). The performance in the train-test-only and external-only models did not vary widely from the overall metrics and can be found in the supplementary tables (Supp Table 2 and 3).

Table 4.

Summary of the performance metrics

Author	Set (model)	AUROC	Accuracy	Recall	Precision	Specificity	F1 score
Ibrahim et al. (2017)	Train-test (SVM)	0.86 ± 0.08	0.86	0.91	0.83	0.80	0.87
	External (SVM)	0.88 ± 0.16	0.88	1.0	0.87	0.00	0.93
Babajani-Feremi et al. (2018)	Train-test (NBC)	0.89 ± 0.07	0.87	0.80	-	0.89	-
Brázdil et al. (2019)	Train-test (SVM)	0.77 ± 0.06	0.75	0.63	-	0.92	-
	Train-test (LDA)	0.68 ± 0.07	0.65	0.49	-	0.88	-
	`External (SVM)	0.79 ± 0.10	0.77	0.58	-	1.0	-
	External (LDA)	0.48 ± 0.13	0.45	0.17	-	0.80	-
Mithani et al. (2019)	Train-test (SVM)	0.93 ± 0.03	0.89	0.96	0.89	0.72	0.92
Mithani et al. (2020)	Train-test (SVM)	0.93 ± 0.04	0.89	0.67	-	0.95	-
	External (SVM)	0.66 ± 0.16	0.67	0.83	-	0.50	-
Fang et al. (2021)	Train-test (RF 1)	-	0.75	0.71	0.80	-	0.75
	Train-test (RF 2)	-	0.65	0.70	0.66	-	0.68
	Train-test (SVM 1)	-	0.73	0.86	0.80	-	0.78
	Train-test (SVM 2)	-	0.69	0.76	0.74	-	0.69
Hao et al. (2021)	Train-test (SVM)	0.75 ± 0.10	0.67	0.70	-	-	-
Ma et al. (2022)	Train-test (SVM)	0.77 ± 0.06	0.76	0.73	0.81	0.79	0.76
	External (SVM)	-	0.61	-	-	-	-
Chen et al. (2023)	Train-test (SVM)	0.88 ± 0.07	0.81	0.91	0.79	0.84	0.86
	Train-test (XGB)	0.85 ± 0.15	-	0.88	-	0.79	-
	External (SVM)	0.91 ± 0.12	0.71	0.87	1.0	0.79	1.0
	External (XGB)	0.90 ± 0.06	0.79	0.92	-	0.74	-
Berger et al. (2023)	Train-test (SVM)	0.82 ± 0.06	0.80	0.75	0.86	0.86	0.80
Cheng et al. (2024)	Train-test (SVM)	0.84 ± 0.05	0.81	0.92	0.79	0.67	0.85
Berger et al. (2024)	Train-test (SVM)	0.88 ± 0.08	0.94	1.0	0.92	0.83	0.96
Overall (mean ± SD)*		0.84 ± 0.09	0.76 ± 0.11	0.77 ± 0.19	0.83 ± 0.08	0.76 ± 0.22	0.83 ± 0.10

AUROC Area under the receiver operating characteristic curve, NBC Naïve bayes classifier, LDA Linear discriminant analysis, RF Random forest, SD Standard deviation, SVM Support vector machine

^*AUROC expressed as weighted AUROC mean and se

Discussion

Developing an accurate model to predict response to neuromodulation therapies can be beneficial for patients with DRE. Currently, response to these therapies is varying, normally ranging between 30% in low-risk of bias papers and up to 70% in high bias risk publications (Ryvlin et al. 2021), with brain-invasive therapies, such as RNS, tending to have higher rates of response than non-brain-invasive therapies such as VNS (70% vs 50% in previous studies) (Alcala-Zermeno et al. 2022; Khan et al. 2022; Englot et al. 2016). In addition to the low responder rates, all these therapies have intrinsic risks, with adverse events reported in up to one-third of the patients (Ryvlin et al. 2021; Lim et al. 2024; Heck et al. 2014). Moreover, these procedures generate important costs for the patients, with an average of $88,000 per patient for VNS and around $100,000 for DBS and RNS (Vincent et al. 2022), highlighting the importance of identifying responders, preventing unnecessary exposure to risk of surgical complications, and costs to potentially unresponsive patients.

A critical focus when building reliable predictive models for neuromodulation outcomes is considering the complexity of these patients. DRE is heterogeneous by nature, with multiple pathophysiological pathways, and usually caused by a combination of different mechanisms like pharmacokinetic alterations, genetic polymorphisms and mutations, altered cellular ASM’s targets, inadequate remodeling and plasticity in neuronal networks, and increased inflammation; among many others (Pérez-Pérez et al. 2021; Tang et al. 2017). Taking this into consideration, multimodal approaches could potentially encompass this complexity in a better way than unimodal ones. However, acquiring and training models with a fully multimodal approach is complex. In our review, the analyzed studies normally leveraged clinical data records and an additional type of data, that included electrocardiograms (Fang et al. 2021), electroencephalograms (Ma et al. 2022; Chen et al. 2023; Cheng et al. 2024), and connectomics via diffusion tract imaging (Mithani et al. 2019), but to our knowledge no previous study have incorporated a truly multimodal approach with easily accessible studies, as some diagnostic modalities like magnetoencephalogram and diffusion tract imaging are not easily accessible in the clinical practice, limiting their applicability.

Most studies examined functional and structural connectivity parameters as predictors, and several of these have been proposed as potential biomarkers of VNS response. In particular, increased connectivity has been associated with better clinical outcomes after VNS (Germany et al. 2023; Danthine et al. 2025), possibly related to the desynchronization effect of VNS in the brain (Vespa et al. 2021). Other focus was on heart rate variability (HRV) and laryngeal motor evoked potentials, which reflect intact vagal function through the parasympathetic autonomic nervous system and adequate transmission of signals from the laryngeal nerves to higher cortical areas (Ottaviani and Macefield 2022; Neuhuber and Berthoud 2021). These three approaches have been explored in multiple studies, with several iterations depending on the modality used. Connectivity can be assessed with EEG, MEG, and fMRI, and these modalities provide the most direct information about brain network organization. In parallel, HRV through the use of ECG and laryngeal motor evoked potentials through EEG have been used to evaluate vagal integrity, function, and brain–body communication.

Interestingly, the AUROC was on average higher in models trained only in pediatric cohorts, compared to adult-only or mixed populations. One of the models developed by Brazdil (Brázdil et al. 2019) had an AUROC of 0.48, which skews the results. Additionally, population differences between adults and children may pose challenges for ML models. In adults, epilepsy often arises from more identifiable and heterogeneous etiologies, encompassing both functional and structural causes (e.g., neoplasms and cerebrovascular disease). In contrast, pediatric epilepsy more commonly involves age-dependent syndromes and genetic etiologies (Babunovska et al. 2021; Vikin et al. 2025).

Moreover, these studies relied on relatively small, single-center or dual-center cohorts. These limited sample sizes restrict statistical power, are prone to overfitting and may reduce performance in different populations. While the features used in these models are diverse and sometimes not easily accessible, preliminary studies demonstrate the feasibility of machine learning approaches for predicting treatment or neuromodulation response, even when based on heterogeneous and multimodal data, based on the performance on the external validation cohorts, however, there is an impending need for additional large-scale, multicenter studies, which are essential to refine these models, identify the most clinically relevant predictors, and make them generalizable for the daily clinical practice.

Conclusion

Multimodal ML presents as a valuable tool for current neurology, as it allows the leverage of different types of data to more accurately predict different outcomes. In the present study we analyzed the performances of different multimodal ML models to predict response to neuromodulation techniques (mainly VNS) as treatment for patients diagnosed with DRE. The overall performance of the models was good, with a pooled AUROC of 0.84 (0.79–0.88), which did not differ significantly when comparing between populations. However, we acknowledge that the risk of bias was high or unclear in more than half of the studies and the studies included small cohorts, which entails their own risks during the training of these architectures.

Abbreviations

ASM

Antiseizure Medication

AUROC

Area Under the Receiver Operating Characteristic Curve

CNS

Central Nervous System

ctDCS

Cathodal Transcranial Direct Current Stimulation

CV

Cross-validation

DBS

Deep Brain Stimulation

DL

Deep Learning

DRE

Drug-resistant epilepsy

ILAE

International League Against Epilepsy

LDA

Linear discriminant analysis

LOO

Leave-one-out

MESH

Medical Subject Headings

ML

Machine Learning

MRI

Magnetic Resonance Imaging

NBC

Naïve bayes classifier

PRISMA

Preferred Reporting Items for Systematic reviews and Meta-Analyses

PROBAST

Prediction Model Risk of Bias Assessment Tool

RCS

Responsive Cortical Stimulation

RF

Random forest

RNS

Responsive Neurostimulation

RScS

Responsive Subcortical Stimulation5

SE

Standard error

SVM

Support vector machine

VN

Vagus Nerve

VNS

Vagus Nerve Stimulation

Authors’ contributions

**A.Q.V.** and **F.F** conducted the screening of articles, performed data extraction, and carried out the statistical and meta-analytic analyses. **J.M** conducted the initial literature search, compiled the preliminary database of studies, provided expert guidance as the research librarian, and performed a specialized database search. **T.Z.** supervised the overall project, resolved conflicts during screening and data extraction, and contributed to the theoretical framework and interpretation of the findings.

Funding

No funding was required for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.