Abstracts

Rationale: Seizures of around one-third of all epileptic patients cannot be controlled by antiepileptic drugs. Epilepsy surgery is the main alternative, but pre-surgical workups do not ensure favorable outcomes. Seizures recur in about half of the operated patients, even after invasive explorations (Mohan et al.; Plos 1 2018 16;13). Stereo-EEG (SEEG) allows recording neural activity with high spatio-temporal resolution and it is thus a valuable resource for defining the epileptogenic network (EN). SEEG quantification is however non-standardized, largely due to the lack of consensus regarding the EN concept (Steinbart et al.; Clin Neurophysiol. 2020 131(11):2682-2690). Most computational approaches rely on mathematical descriptions of local field potentials during seizures. On the other hand, connectivity approaches try to define EN connectivity as compared to non-epileptogenic networks (NN). No strategy so far has provided consistent results in terms of surgical outcome prediction. We propose a model for EN and NN connectivity dynamics through a game theory concept. Each SEEG channel is a network node that either plays for, or against the EN. We analyze the transition from non-seizure to seizure to assign nodes to either network.

Methods: As the game theory suggests, players strive to maximize their payoffs, which in this case would be maintaining the default state vs. switching to ictal state. Connectivity measures between each pair of nodes are the features that determine states. A support vector machine classifies between default and ictal epochs, and random data splits are applied to simulate different game scenarios (Fig. 1). The probability of epileptogenic nodes winning is scored using the minimax algorithm, a decision rule often used in game theory. Since the transition is to seizure, the nodes with highest scores were selected as EN. We validate the framework on a chronological cohort of 21 drug-resistant epilepsy patients, with the only inclusion criterion of a 3-year follow-up.

Results: A surgical outcome prediction accuracy of 93% was achieved, which is the best to our knowledge. Several time intervals prior to and during seizures were tested. The EN could not be inferred from the seizure event itself. Instead, an optimal time interval for EN definition was at the transition from pre-seizure to seizure (Fig. 2).

Conclusions: So far ictal activity is considered the best source to assess EN topology and surgery outcomes. In the proposed model, the transition between non-seizure and seizure, when the EN prevails over the NN, is the best time to distinguish the networks (Fig. 2). After seizure onset, NN connectivity changes as epileptogenic activity propagates, while before onset, EN connectivity could be influenced by NN. In future works, the mechanisms of this transition will be addressed. Our work not only reinforces the network concept of epileptogenicity but also provides a tool for EN definition, which may find applications in cognitive neuroscience.

Funding: Please list any funding that was received in support of this abstract.: No funding was received in support of this abstract.