From Computational Models to Clinical Decision-Making in Epilepsy: Bridging the Gap Between Theory and Practice.

Epilepsy is one of the most heterogeneous and unpredictable neurological disorders, characterized by complex brain dynamics, significant interindividual variability, and a therapeutic response that is often difficult to predict. Despite major advances in neuroimaging, electrophysiology, and antiepileptic treatments, clinical decision-making still faces major challenges, including precisely localizing the epileptogenic zone, predicting seizures, identifying drug-resistant patients, and optimizing personalized therapeutic strategies. Over the past two decades, computational models have gradually emerged as promising tools for addressing these challenges. Early approaches relied primarily on mechanistic biophysical models aimed at reproducing the neural dynamics and mechanisms of hyperexcitability involved in epileptogenesis. More recently, the rise of artificial intelligence and deep learning has facilitated the development of predictive models capable of leveraging large-scale multimodal data. Finally, the emergence of digital brain twins and adaptive closed-loop systems now paves the way for personalized computational medicine integrating simulation, prediction, and real-time therapeutic intervention. This review provides a critical analysis of the major international initiatives developed between 2004 and 2026, including the EPICLUSTER/EBRAINS, EPINOSTICS, ENIGMA-Epilepsy, NeuroPace RNS, SeizeIT, EpiBioS4Rx, and NeuroVista projects, as well as several translational platforms based on artificial intelligence and biophysical modeling. We evaluate the actual contribution of these models to clinical decision-making in epilepsy, their translational contributions, their methodological limitations, and the challenges associated with their integration into hospital settings. We argue that the future of computational epileptology does not rest on the opposition between artificial intelligence and mechanistic models, but on their convergence within explainable, personalized, and clinically interpretable hybrid architectures.