Seizure dynamotype classification using non-invasive recordings

Summary Objective Recently, a seizure classification approach derived from complex systems and nonlinear dynamics has been suggested, termed the “taxonomy of seizure dynamotypes.” This framework is based on modeling the dynamical process of the transition in and out of a seizure. It has been examined in computational and animal models in-vitro and recently in human intracranial data. However, its applicability and value in surface EEG remain unclear. This study examined the applicability of dynamotype classification to seizure information extracted from surface EEG and tested how it relates to clinical factors. Methods Surface EEG recordings from 1,215 seizures were analyzed. We used an automated pre-processing pipeline, resulting in independent components (ICs) for each seizure. Subsequently, we visually identified ICs with clear seizure information and classified them based on the suggested taxonomy. To examine the possibility of automatic classification, we applied a random forest classifier combined with EEG features and evaluated its performance in identifying seizure-related ICs and classifying dynamical types. Lastly, we used a Bayesian estimator to examine the likelihood of the different dynamical types under various clinical conditions. Results We found an apparent onset and offset bifurcation in 49.5% and 40.3% of seizures, respectively. Bifurcation prevalence aligns with that previously reported using intracranial data and computational modeling. The automated classifiers, evaluated with a leave-one-patient-out paradigm, provided good performance. In addition, bifurcation prevalence differed between vigilance states and seizure classes. Significance We demonstrated a method to extract seizure information and classify dynamotypes in non-invasive recordings with a visual as well as an automated framework. Extending this classification to a larger scale and a broader population may provide further insights into seizure dynamics. Key Points Box We characterized the dynamical types of transitions at seizure onset and offset based on seizure information extracted from surface EEG. We classified the dynamical types (dynamotypes) in 49.5% and 40.3% of seizure onsets and offsets, respectively. The dynamotype distribution in surface EEG data aligns with previous findings from intracranial EEG and theoretical expectations. The likelihood of the dynamical type of a seizure exhibits differences across clinical seizure classes and vigilance states. Automated detection and classification of seizure bifurcations are possible using relevant features and pre-existing tools.