Machine learning-based models for predicting glioma-associated epilepsy: a systematic review and meta-analysis

Abstract

Background

Glioma-associated epilepsy (GAE) is a common and disabling complication in glioma patients. Predicting seizures in this population is challenging due to complex tumor–host interactions. With recent advancements in machine learning (ML) models, these models can incorporate high-dimensional datasets and detect subtle patterns. This systematic review and meta-analysis aimed to evaluate the predictive performance of ML-based models for predicting GAE.

Methods

A comprehensive review was performed following PRISMA guidelines in four databases (PubMed, Embase, Scopus, and Web of Science) on May 23, 2025. Studies developing ML-based models for GAE prediction were included. Pooled estimates for area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and diagnostic odds ratio (DOR) were calculated.

Results

Thirteen studies with 3,253 patients were included. Pooled AUC was 0.87 (95% CI: 0.83–0.91), and ACC was 0.82 (95% CI: 0.76–0.88). The pooled SEN was 0.77 (95% CI: 0.64–0.87), SPE was 0.93 (95% CI: 0.86–0.96), and DOR was 40.1 (95% CI: 17.1–94.0). The Summary Receiver Operating Characteristic (SROC) curve demonstrated a false positive rate of 0.09.

Conclusion

ML-based models demonstrate encouraging diagnostic performance in predicting GAE. Incorporating these models into daily clinical practice can help physicians with risk stratification and the identification of high-risk individuals, thereby optimizing therapeutic strategies and enhancing patient outcomes. Before implementing these models in real-time clinical practice, several limitations, including a lack of standardized protocols, considerable heterogeneity among models, and a lack of external validation, should be addressed.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04035-4.

Introduction

Glioma-associated epilepsy (GAE) is a frequent and disabling presentation in individuals with glioma [1, 2]. Prior studies have demonstrated that the risk of GAE occurrence is inversely proportional to the World Health Organization (WHO) lesion grade, with lower-grade lesions being associated with a higher risk of GAE [1–5]. The risk of GAE development in low-grade gliomas (LGGs) has been reported to be 60–85%, whereas the risk for high-grade gliomas (HGGs) is 30–50% [1–5]. GAEs in individuals with glioma can significantly impact the quality of life and complicate the decision-making process in this population [1–5]. Despite significant advancements in neuroimaging and electrophysiology, predicting the development of epilepsy in glioma patients remains challenging, as it is linked to various factors and their complex interactions, such as the biology of the lesion, location, and the patient’s predisposition.

With recent advancements, machine learning (ML)-based models have demonstrated a promising predictive performance in various neurological conditions [6–9]. Through the use of high-dimensional datasets comprising radiological, histopathological, and clinical data, these models can uncover patterns that conventional approaches are unable to identify [6–9]. Several studies have investigated the predictive performance of ML-based models in predicting GAE [10–22]. These models have demonstrated encouraging predictive performance for predicting epilepsy in glioma patients [10–22]. However, these models vary considerably in design, model type, data input, and validation methods, complicating direct comparisons and generalizability.

This systematic review and meta-analysis aim to evaluate the predictive performance of ML-based models in GAE prediction by synthesizing the available data. We strive to assess the methodological quality of current models and quantify model performance through pooled estimates of area under the curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE). Through addressing the strengths and limitations of these models, we aim to guide physicians in the decision-making process for glioma patients and facilitate the integration of these models into clinical practice.

Materials and methods

Objective

We aimed to evaluate the predictive performance of the ML-based models in GAE prediction following the “Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)” guidelines [23].

Search strategy

A comprehensive literature search was performed in the PubMed, Embase, Scopus, and Web of Science databases on May 23, 2025. The search used the terms “Machine learning,” “Deep learning,” “Artificial intelligence,” “Radiomics,” “Glioma,” “Epilepsy,” and “Seizure” along with their equivalents (Supplementary Table S1).

Eligibility criteria

The PICO framework is provided in Supplementary Table S2. Inclusion criteria:

• Studies that developed predictive models for glioma patients.

• Studies that developed ML, deep learning (DL), or neural network (NN) models to predict GAE.

• Studies that reported at least one of the AUC, ACC, SEN, or SPE.

Exclusion criteria:

• Reviews, book chapters, conference abstracts, preprints, and editorials.

• Studies that used conventional models without ML, DL, or NN application.

• Studies that lack AUC, ACC, SEN, or SPE specifically for GAE prediction.

• Studies published in a language other than English.

Study selection process, data extraction, and risk of bias assessment

Two independent reviewers performed the title and abstract screening using Covidence, and a third reviewer handled the disagreements. Then, two independent reviewers conducted the full-text screening, and a third reviewer addressed any conflicts. The included studies underwent data extraction by two independent reviewers, and any disagreements were resolved by a third reviewer. The extracted variables are provided in the Supplementary Table S3. The Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) was applied by two independent reviewers to assess the risk of bias (RoB), and any discrepancies were handled by a third author [24].

Statistical analysis

The R program (version 4.4.2), utilizing the “meta,” “metafor,” and “mada” packages, was employed for statistical analysis. For each model, we extracted the highest reported AUC for GAE prediction, prioritizing values from external or independent test sets when available. Otherwise, the highest value from internal validation or training sets was applied. Pooled estimates of AUC and ACC were calculated using inverse-variance random-effects meta-analysis, and the 95% confidence intervals (CIs) for ACC and AUC were determined with the Wilson score interval and the Hanley and McNeil method when the 95% CI was not available. The pooled estimates for SEN, SPE, and diagnostic odds ratio (DOR) were calculated using logit-transformed proportions. Leave-one-out sensitivity analysis assessed the robustness of the findings. Significant moderators were identified based on p-values < 0.05. To unify the parameters, if a variable was reported as a median, it was converted to a mean using the method described by Luo et al. (46). The definitions of the outcomes are provided in Supplementary Table S4.

Results

Study selection process

The search yielded 769 studies across four databases (Fig. 1). Of these, 271 were duplicates and were removed, resulting in 498 studies enrolled for title and abstract screening. During this screening, 456 were excluded for not meeting the eligibility criteria, and 42 proceeded to full-text screening. Of these 42, 29 were excluded, 21 because they were not ML models, and eight for being abstracts. Ultimately, 13 studies were included for data extraction [10–22].

Fig. 1.

PRISMA flowchart of the included studies

Risk of bias assessment

The RoB assessment of the included studies, conducted with the QUADAS-2 tool, showed that most studies had an overall ‘Low’ or ‘Some concerns’ RoB (Supplementary Fig. S1). The RoB mainly stemmed from D2 (Index Test) and D3 (Reference Standard) domains due to an unclear blinding process or the use of non-pre-specified thresholds in model evaluation. These issues raise concerns about the potential for bias in the development and interpretation of ML models. Overall, the QUADAS-2 results underscore the need for cautious interpretation of the current study’s findings.

Baseline characteristics and performance outcomes

Thirteen models, including 3,253 glioma patients, participated in our study (Table 1). The prevalence of GAE was 37.4% (1,216/3,253) in the cohort. The male and female groups comprised 51.6% (1,066/2,066) and 48.4% (1,000/2,066) of the cohort. The World Health Organization (WHO) grade IV (35.8%, 762/2,130) was the most common grade, followed by grade III (33.0%, 702/2,130), grade II (29.7%, 633/2,130), and grade I (1.5%, 33/2,130). The machine learning (ML) models accounted for 92.3% (12/13) of the models, while deep learning (DL) models made up 7.7% (1/13). The validation methods included train-test split in 69.2% (9/13) of the models, leave-one-out cross-validation (LOOCV) in 23.1% (3/13), and 10-fold cross-validation (CV) in 7.7% (1/13). The input data types were radiomics and clinical in 53.8% (7/13), radiomics alone in 23.1% (3/13), genomics and clinical in 7.7% (1/13), image alone in 7.7% (1/13), and genomics alone in 7.7% (1/13). The support vector machine (SVM) was the most frequently used algorithm, accounting for 38.5% (5/13). The AUC ranged from 0.792 to 1, and the ACC ranged from 0.75 to 1. The sensitivity (SEN) ranged from 0.156 to 1, and the specificity (SPE) ranged from 0.815 to 1.

Meta-analysis of outcomes

Twelve studies were included in the meta-analysis of AUC for GAE prediction in glioma (Fig. 2A). The meta-analysis showed a pooled AUC of 0.87 (95% CI: 0.83–0.91). Ten studies were included in the meta-analysis of ACC (Fig. 2B). The meta-analysis showed a pooled 0.82 (95% CI: 0.76–0.88). Seven studies provided both SEN and SPE and were included in the meta-analyses of SNE, SPE, and DOR. The results revealed a pooled SEN of 0.77, SPE of 0.93 (95% CI: 0.86–0.96), and DOR of 40.1 (95% CI: 17.1–94) (Fig. 3A–C). The Summary Receiver Operating Characteristic (SROC) curve displayed an AUC of 0.763, SEN of 0.759, and a false positive rate of 0.09 (Fig. 3D).

Fig. 2.

Meta-analysis forest plots showing model performance for predicting GAE in gliomas using ML models. A Pooled AUC, B Pooled ACC

Fig. 3.

Meta-analysis forest plots showing model performance for predicting GAE in gliomas using ML models. A Pooled sensitivity, B Pooled specificity, C Pooled DOR, D SROC curve

Sensitivity analysis

The sensitivity analysis for the AUC and ACC showed that the estimates remained consistent, which indicates the robustness of the findings (Supplementary Fig. S2A-B). The sensitivity analysis for the SEN, SPE, and DOR also demonstrated that none of the included studies had a significant impact on the pooled estimates, and the results were robust (Supplementary Fig. S2C–E).

Discussion

This systematic review and meta-analysis evaluated the role of ML-based models in the prediction of GAE. Our findings revealed that ML-based models demonstrated promising predictive performance for GAE prediction, with considerable AUC and ACC values. These findings support the growing role of ML in neuro-oncology, especially in the context of seizure risk stratification among glioma individuals.

Development of seizures in glioma individuals primarily stems from the lesion’s direct and indirect impact on the brain’s cortical excitability and neural networks [1, 25–27]. Gliomas, especially LGGs and cortical-subcortical lesions, typically infiltrate or irritate surrounding healthy tissue, resulting in disrupted neuronal architecture and abnormal electrical activity [1, 25–27]. Tumor-induced alterations, including peritumoral inflammation, blood-brain barrier disruption, glutamate release, altered expression of ion channels, and gliosis, contribute to hyperexcitability [1, 25–27]. In addition, molecular biomarkers, including isocitrate dehydrogenase mutations and 1p/19q codeletions, have been linked with higher susceptibility to seizure development, further connecting tumor biology with epileptogenesis [1, 25–27].

The diagnostic performance was consistently promising, especially in those models that integrated radiomic, clinical, and molecular features. In the study by Gao et al., a model combining radiomics with clinical data based on T2-weighted Fluid-Attenuated Inversion Recovery (T2-FLAIR) achieved an AUC of 0.886 in the training cohort and 0.836 in the test cohort, outperforming the radiomics-only model [21]. George et al. developed an XGBoost model using multiparametric magnetic resonance imaging (MRI) sequences, leading to an accuracy of 0.81 [10]. Zhong et al. also highlighted the added value of multiomics data by integrating radiomic and urinary proteomic markers into a decision tree model, achieving AUCs of 0.897 and 0.874 in the training and validation cohorts, respectively [12].

The results of this meta-analysis provide the first quantitative benchmark for the diagnostic accuracy of ML-based models in GAE. A pooled AUC of 0.87 and specificity of 0.93 indicate that these models can accurately identify patients at high risk of postoperative or perioperative seizures. In practice, this information could be integrated into preoperative risk stratification: neurosurgeons could tailor antiepileptic prophylaxis, plan extent of resection, and counsel patients regarding seizure risk. For neuro-oncologists, model-derived risk maps may guide the timing of adjuvant therapy and facilitate early intervention for seizure control. For radiologists, radiomic-based predictors may highlight epileptogenic tumor subregions, informing targeted resection or SRS planning. Thus, beyond technical validation, our synthesis provides evidence that ML-driven prediction can enhance precision care in glioma management. Despite the proposal of multiple ML models for GAE prediction, their heterogeneous methodologies and single-center validations have impeded their practical application. Our meta-analysis aggregates these fragmented results, establishing the first pooled performance estimates that clarify the reliability and generalizability of ML-based epilepsy prediction in glioma.

However, several issues need to be addressed before these models are integrated into daily clinical use [28–32]. These models are developed using various input types, such as radiomics, genomics, and clinical data, which can limit their generalizability. Most current models are based on single-center cohorts and lack external validation, which can lead to overfitting [28–32]. Additionally, interpreting the results of these models can be challenging, especially for those based on DL methods, which are more complex than other models [28–32]. Additionally, the real-time application of these models necessitates their integration into existing systems, such as EPIC and PACS, with a user-friendly interface. These issues should be addressed before applying these models effectively in clinical practice.

Heterogeneity across the included studies poses a significant challenge in the existing ML-based models for GAE prediction. This problem arises from various sources during the development of these models. The cohorts of each model differ significantly in patient characteristics, such as age and gender, as well as lesion-related features, including WHO grade, prior treatments, and the time point of inclusion in the cohort. Another source is the data input used by the models; some utilize radiomics, others genomics, and some combine these with clinical data. Even among those who use radiomics, different MRI sequences are employed, leading to considerable heterogeneity. Without standardized protocols and definitions, it becomes difficult to determine whether these differences in performance reflect actual variations in predictive power or are artifacts of data and process inconsistencies. This issue should be addressed before these models are routinely incorporated into clinical practice.

In this study, we sought to mitigate heterogeneity by employing a random-effects meta-analytic framework, which assumes that true model performance may vary across studies. This approach, coupled with leave-one-out sensitivity analyses, ensured that no single study disproportionately influenced the pooled results. Although variability in input features, algorithms, and validation methods remains inherent to ML-based research, our findings represent an average performance estimate across diverse modeling strategies, reflecting real-world heterogeneity. Moving forward, standardization of feature engineering, model reporting, and external multicenter validation will be crucial to enhancing reproducibility and facilitating clinical translation.

This study is associated with several limitations that have to be acknowledged. Most of the included studies used single-center and retrospective cohorts to develop models that may introduce overfitting and limit the generalizability of the findings. Another limitation is the presence of high heterogeneity, which stems from variations in patient characteristics, lesion features, data types (radiomics, genomics, clinical), and imaging protocols, making direct comparisons complicated. Furthermore, the application of various ML algorithms and inconsistent reporting of performance metrics, such as CIs for AUC and accuracy, restricts rigorous benchmarking across models. Furthermore, none of the included studies employed external validation, and all models were internally validated using methods such as train–test split, k-fold cross-validation, or leave-one-out cross-validation. This lack of external validation limits the generalizability of the reported model performance and raises the potential for overfitting to single-institution datasets. Another limitation is that the included studies encompassed both adult and pediatric glioma populations. Given the distinct pathological and molecular characteristics between these groups, combining them may introduce heterogeneity and limit the generalizability of pooled model performance estimates.

Future studies should focus on using large, prospective, multi-center datasets to validate ML-based models for GAE prediction. They should also work on developing a standardized protocol for creating ML-based models to ensure high reliability and broad applicability. Additionally, future research should include external validation to reduce overfitting and improve the generalizability of these models.

Conclusion

This systematic review and meta-analysis demonstrated that ML-based models are associated with encouraging diagnostic performance in predicting GAE. Incorporation of these models into daily clinical routine can help physicians with risk stratification and the identification of high-risk individuals, thereby optimizing therapeutic strategies, enhancing patient outcomes, and improving the quality of life for glioma patients. Before implementing these models in real-time clinical practice, several limitations, including a lack of standardized protocols, considerable heterogeneity among models, and a lack of external validation, should be addressed.

Author contributions

Conceptualization: B.H., M.H.; Methodology: B.H., I.M., S.T.; Literature Search: B.H., A.E., P.S.; Data Extraction: D.N., E.B.H., A.E.; RoB Assessment: D.N., E.B.H., A.E.; Statistical Analysis: B.H.; Writing – Original Draft: B.H., S.T., F.Gh.; Writing – Review & Editing: B.H., M.H.; Supervision: B.H.; Project Administration: B.H.

Funding

The authors declare that no funding source was used.

Data availability

The data supporting this study’s findings are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study is deemed exempt from receiving ethical approval.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.