Use of artificial intelligence in magnetic resonance imaging across the epileptic patient's journey: A meta‐analysis of four clinical applications

OBJECTIVE: The application of artificial intelligence/machine learning (AI/ML) to magnetic resonance imaging (MRI) promises to enhance and support clinical decision-making in epilepsy. However, there currently lacks an appropriate assessment of clinical utility and study rigor of current AI/ML-driven models that are targeted toward supporting decision-making within the clinical workup in epilepsy. METHODS: We systematically reviewed and examined the ability of current AI/ML-driven models in MRI across four main applications within the clinical workup(s) for epilepsy: (1) diagnosis, (2) temporal lobe epilepsy lateralization, (3) lesion (focal cortical dysplasia) localization, and (4) postsurgical outcome prediction. We additionally assessed the risk of bias for each study model. Studies that employed AI/ML classification models trained on any MRI modality or sequence type were selected for qualitative assessment; those reporting accuracy rates were subsequently included in the meta-analysis. RESULTS: Of 3227 searched articles, we identified 159 studies (n = 26 732 participants) for qualitative evaluation and 127 studies (n = 20 456) for inclusion in the meta-analysis. Our results reveal that AI/ML on MRI could accurately distinguish epilepsy patients from healthy controls (overall accuracy = .87, 95% confidence interval CI = .85-.89), lateralize temporal lobe epilepsy (.90, 95% CI = .87-.93), localize epileptogenic lesions (.82, 95% CI = .74-.87), and predict postsurgical seizure freedom (.83, 95% CI = .78-.87). However, systematic assessment indicated a very high risk of bias in the literature, suggestive of overly optimistic performance estimates. SIGNIFICANCE: Although our results support overall high accuracy of AI/ML models in epilepsy diagnostics and prognostics, the literature remains susceptible to bias in participant recruitment and validation methods. Furthermore, most models were limited by study architecture that demands strict adherence to nonstandard, highly specific data acquisition and processing protocols that cannot be easily deployed for clinical implementation. We encourage closer interdisciplinary collaboration between clinical and scientific groups to improve validation studies, and outline suggested recommendations for future study design, analysis, and reporting.