Abstracts

Rationale: Epilepsy, one of the most common neurological syndromes, is increasingly recognized as involving complex brain network dynamics, making its understanding and treatment a unique challenge. The diversity of electrocorticographical (ECoG) patterns observed during seizures, particularly at seizure onset, naturally leads to the question of whether different ECoG patterns result from different underlying mechanisms. Previous work has investigated seizure onset patterns in animal models of epileptic disorders as well as in humans. However, there is yet no consensus on (1) how to cluster different ECoG patterns, and (2) whether these patterns are triggered by different mechanisms. Here we address these issues by employing unsupervised machine learning algorithms to cluster seizure onset patterns from human ECoG recordings. We hypothesize that seizures belonging to each cluster share features that can provide insights into their mechanism of generation. Methods: We pursued two parallel approaches to cluster the onset patterns of 256 seizures recorded from 27 patients with intractable epilepsy who underwent presurgical evaluation. In one approach, a set of well-defined features was computed from the waveforms, spectra and entropy measures of ECoG signals. In the second approach, features were automatically extracted using a deep convolutional denoising autoencoder with separate hidden nodes for waveform and spectral patterns. In both approaches, the difference in features before and after seizure onset was fed to a hierarchical clustering algorithm. We evaluated how many types of onset patterns best describe the data. We also studied which ECoG features account for clustering. Results: We found with both approaches that a small number of clusters account for the majority of seizure onset patterns observed in our population of patients. The feature-based analysis also revealed that, although certain EEG features change similarly for all seizure onsets, other vary greatly across clusters. Overall, the fact that we found using either approach distinct clusters combining seizures from different patients suggests a limited number of possible type of ECoG dynamics, and therefore of mechanisms, during seizure initiation. Conclusions: We propose that being able to reliably and objectively cluster the patterns at seizure onset using unsupervised machine learning gives us insights into possible mechanisms of seizure generation. The difference across patterns may arise from differences in neuronal involvement during ictogenesis. Understanding the mechanisms may provide us with novel approaches to improve patient care of medically refractory epilepsy, such as refinements in surgical target identification, surgical outcome prediction and alternative seizure control strategies. Funding: Sun, Westover: NIH-NINDS 1K23NS090900, 1R01NS102190Cash, Martinet: R01-NS062092, 1K24NS088568-01A1, R01-NS079533, R01_NS072023Salami: Postdoctoral fellowship from Fonds de Recherche Santé Québec (FRSQ)