Abstracts

Rationale: Epileptic seizures have been at the core of many studies focusing on the various aspects of their generation, maintenance and propagation. Ictal events are cyclic but their phase is not constant. Despite this simple evidence, at present there is no unified way to describe the mechanisms underlying seizure preparation and termination, especially not at a network level. Over the last decades, a number of analytical approaches have highlighted interesting features of the epileptogenic network, ranging from electrical (local) to anatomical and functional connectivity (global) patterns, allowing prediction (models) or warning (applications) about the upcoming seizure. Despite the many efforts, the most clinically relevant solutions analyze the dynamics of brain areas separately to derive an anatomical description of the epileptogenic network extent. Methods: In our work we show how different brain regions explored with invasive EEG influence each other through whole seizure clusters. For this purpose we developed a machine learning algorithm designed to highlight prediction error asymmetries between parallel sources, and processed a feature of the EEG mainly dependent on synaptic activity. We analysed more than 81 consecutive hours of Stereo-EEG (SEEG) recordings of refractory epilepsy patients that presented clusters of seizures. When trying to predict the activity dynamics of two different brain regions we hypothesised to find either similar, or different prediction error yields. When the latter occurred we considered that the region with higher yields influenced the second since it comprehended most parts of both regions dynamics. Results: Only when the influence of some nodes overcomes the rest or, in other words, when the network is divided between masters and slaves, ictal events occur. Seizures reset the cycle, allowing the same or other nodes, at times neighboring, other times functionally related, to restart the loop. In the figure we show the weighted (hub nodes selected by time linkage vs. all others) influence levels compared between events and patients. We show all data (curve graph, top) using 6 bins for cluster initiation and termination, and 4 bins for the seizure cycle, using mean (line) and standard deviation (halo). We use 2 bins (postictal and preictal) for statistical comparisons (bars, bottom). Colours reflect seizure occurrence likelihood (red, higher; blue, lower). Error bars represent the standard deviation of master nodes weighted influence. *, p < 0.05; **, p < 0.0005. Moreover, through the anatomical position of the master nodes we could calculate the extent of the epileptogenic network and relate it to the surgery outcome of our series of patients who underwent a minimal resection. Conclusions: Despite the evidence that brain regions are crucial for focal epileptogenesis, this work highlights how seizures ultimately result from network processes, and proposes a novel method to pinpoint the epileptogenic hubs and quantify the probability of ictal events over long stretches of time. Funding: No funding was received in support of this abstract.