Machine learning models for predicting therapeutic response in drug-resistant epilepsy: A bivariate diagnostic meta-analysis

Epilepsy affects approximately 6.4 per 1000 individuals worldwide. Despite advances in antiseizure medications (ASMs), nearly one-third of patients remain refractory to treatment, meeting criteria for drug-resistant epilepsy (DRE). This meta-analysis evaluated the performance of machine learning (ML) models in predicting therapeutic response in DRE based on ILAE-defined outcomes.

A systematic review and meta-analysis were conducted in accordance with PRISMA-DTA guidelines. Studies applying ML algorithms to predict treatment response in DRE were included. Eligible designs comprised retrospective or prospective observational studies. Outcomes included seizure remission (ILAE Class 1), ≥ 50 % reduction in seizure frequency, treatment failure, and diagnostic accuracy metrics. A bivariate random-effects model was used to pool sensitivity, specificity, and diagnostic odds ratio (DOR). Heterogeneity was assessed (I², χ²), and subgroup and meta-regression analyses were performed. Likelihood ratios, area under the curve (AUC), and Bayesian post-test probabilities were estimated. Publication bias was evaluated with Deeks' test.

Eight studies (n = 1887) met inclusion criteria (Kappa=0.98). Pooled sensitivity and specificity were both 0.84 (95 % CI: 0.76-0.89 and 0.77-0.89, respectively), with an AUC of 0.91. DOR was 27; LR+ , 5.2; LR-, 0.19. Heterogeneity was high (χ²=22.7; p < 0.001). Sensitivity was lower in prospective studies, prognostic models, and ASMs users, and higher with long-term follow-up. Meta-regression identified model type, study design, ASM exposure, and sample adequacy as key moderators.

ML models demonstrate high diagnostic accuracy in predicting therapeutic response in DRE. Findings support their potential clinical utility, provided external validation and methodological standardisation.

CRD 42024628323.