Authors :
Presenting Author: Karla Ivankovic, MSc – Hospital del Mar Research Institute, Barcelona, Spain
Alessandro Principe, MD, PhD – Hospital del Mar Research Institute, Barcelona, Spain
Justo Montoya, MSc – Pompeu Fabra University, Barcelona, Spain
Linus Manubens-Gil, PhD – SEU-Allen Joint Ctr. for Neuron Morphology, Southeast Univ. (SEU), Jiangsu, China
Riccardo Zucca, PhD – Radboud University, Nijmegen, Netherlands
Mara Dierssen, MD, PhD – Centre for Genomic Regulation, Barcelona, Spain
Rodrigo Rocamora, MD, PhD – Hospital del Mar Research Institute, Barcelona, Spain
Rationale:
Seizure generation is inherently competitive, with the epileptogenic network (EN) attempting to propagate seizure activity while other networks try to suppress it. A competition-like interaction between the regulator networks and the EN seems to determine whether a seizure will occur or not. EN dynamics depend on the regulatory inputs from other brain networks. As the regulation ceases, the brain enters a connectivity state critical for the seizure to occur. We hypothesized that the EN nodes present maximal connectivity change from the interictal to the critical state. We describe the dynamics between network nodes as a competition and analyze connectivity changes within a game-theoretical framework. We implemented a game of cards between multiple players, in which the highest card wins a turn. The concept of this game was transferred to estimate the winners among brain nodes, based on their connectivity change.
Methods:
Neural populations recorded by intracranial EEG were network nodes. Node connectivity states were quantified in interictal epochs and 1 minute before seizure, using several connectivity measures. A support vector machine with K-fold cross-validation produced scores used to approximate node connectivity change. The scores were used as cards comprising a deck of each player. Players were random groups of nodes, with size equal to 10% of the resection. The winners were the groups scoring above four standard deviations from the mean win score. Resection overlap ratio between winners and losers was used to predict surgery outcomes. The approach was validated on an internal cohort of 21 consecutive patients (3-year follow-up), and an external dataset of 24 patients (1-year follow-up). External validation used the seizure onset zone (SOZ) as ground truth, as resection was unavailable.Results:
The winners had significantly higher resection overlap than the losers. The most prominent connectivity changes at pre-seizure were between EN and non-EN nodes. The patients’ post-surgical outcomes were perfectly classified in the internal cohort (AUC = 100%). The outcome prediction on the external dataset reached the AUC of 96%. The resection overlap ratio between winners and losers consistently separated the good and poor outcomes at the value of 1.Conclusions:
A machine-learning-based approach to computing the connectivity change is independent of a training cohort, and can be applied directly. Analyzing the connectivity change as a competition between network nodes allows for identifying surgical resections with unprecedented accuracy. With the limitation of using SOZ, the external validation was less accurate, but the sensitivity matched the original analysis. The predictive power became optimal with multiple connectivity measures, suggesting that a combination of connectivity features should be considered to generalize the model across patients. This work provides a tool that may aid surgical decision-making, and adds insight into seizure generation mechanisms, supporting competition-like network dynamics.Funding:
Ajuts destinats a universitats, centres de recerca i fundacions hospitalàries per contractar personal investigador novell per a l’any 2022 (FI-DGR Grant Number: 14802)