Title: Optimizing Virtual Epilepsy Neurostimulation via Literature-Informed Minimax Approaches

Abstract

While computational models hold significant potential for tailoring epilepsy treatments to individual patients, current optimization strategies predominantly seek parameter sets that yield average performance. In the context of neuromodulation, this objective is insufficient, as a protocol that enhances the mean response may still prove detrimental to patients with networks that are least resilient to stimulation. To address this, we introduce a minimax pipeline guided by medical literature, integrating hypothesis extraction from PubMed, simulations using the Virtual Brain (TVB) Epileptor, and black-box optimization directed by large language models.

This optimizer generates proposals for either intrinsic model-control parameters or clinically interpretable external-stimulation protocols. The TVB framework then assesses each proposal within a cohort of sampled virtual patients. The primary objective function is designed to maximize the worst-case reward, calculated as the negative variance of the simulated seizure activity.

In experiments focusing on intrinsic model control, the optimal archived parameter set increased the worst-case reward from -0.5285 to -0.3182, marking a 39.8% improvement over the baseline. Conversely, the search for clinical-style external stimulation yielded a marginal worst-case improvement of 1.7%. Although a 20-patient virtual cohort demonstrated a 55% responder rate and a positive signal within the temporal-lobe subgroup, no aggregate benefit was observed (p=0.9019). This study serves as an in silico proof of concept for designing robust, literature-aware neurostimulation protocols and should not be interpreted as clinical evidence.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC