The role of epidemic spreading in seizure dynamics and epilepsy surgery

Abstract

Epilepsy surgery is the treatment of choice for drug-resistant epilepsy patients, but one in three patients continue to have seizures one year after surgery. In order to improve the chances of good outcomes, computational models of seizure dynamics are being integrated into surgical planning to simulate the effects of the planned surgeries. These modelling frameworks require several conceptual and methodological choices, as well as large amounts of patient-specific data, which hinders their clinical applicability. To address this problem, we considered the patient-specific brain network, derived from magnetoencephalography (MEG) recordings, and a simple epidemic spreading model as the dynamical basis for seizure propagation. This simple model was enough to reproduce the seizure propagation patterns derived from stereo-tactical electroencephalography recordings (SEEG) of all considered patients (N=15), when the patients' resected areas (RA) were used as the origin of epidemic spreading. The model yielded a more accurate fit for the seizure-free (SF, N=11) than the non-SF (NSF) group and, even though the difference between the groups was not significant, the goodness-of-fit distinguished NSF from SF patients with an area under the curve AUC=84.1%. We also explored the definition of a population model that combined data from different patients to fit the model parameters but was still individualized by considering the patient-specific MEG network. Even though the goodness-of-fit decreased compared to the individualized models, the difference between the SF and NSF groups held, and in fact became stronger and significant (p=0.023), and the group classification also improved slightly (AUC=88.6%). Therefore, combining data from different patients may pave the way not only to generalize this framework to patients without SEEG recordings, but also to reduce the risk of over-fitting and improve the stability of the models. Finally, we considered the individualized models to derive alternative hypothesis of the seizure onset zones and to test the surgical strategy in silico for each patient. We found that RA regions were on average more likely to originate the seizures, but that alternative explanations were possible. Virtual resections of the RA when considering these alternative seeds significantly reduced seizure propagation, and to a greater extend for SF than NSF patients (although the difference was not significant). Overall, our findings indicate that spreading models based on the patient-specific MEG network can be used to predict surgical outcomes, with better fit results and greater reduction on seizure spreading linked to higher likelihood of seizure freedom after surgery.