Abstract
Epilepsy, a neurological disorder characterized by recurrent unprovoked seizures, significantly impacts patient quality of life. Current classification methods focus primarily on clinical observations and electroencephalography (EEG) analysis, often overlooking the underlying dynamics driving seizures. This study uses surface EEG data to identify seizure transitions using a dynamical systems-based framework-the taxonomy of seizure dynamotypes-previously examined only in invasive data. We applied principal component and independent component (IC) analysis to surface EEG recordings from 1,177 seizures in 158 patients with focal epilepsy, decomposing the signals into ICs. The ICs were visually labeled for clear seizure transitions and bifurcation morphologies (BifMs), which were then examined using Bayesian multilevel modeling in the context of clinical factors. Our analysis reveals that certain onset bifurcations (saddle node on invariant circle and supercritical Hopf) are more prevalent during wakefulness compared with their reduced rate during nonrapid eye movement (NREM) sleep, particularly NREM3. We discuss the possible implications of our results in the context of modeling approaches and suggest additional avenues to continue this exploration. Furthermore, we demonstrate the feasibility of automating this classification process using machine learning, achieving high performance in identifying seizure-related ICs and classifying interspike interval changes. Our findings suggest that the noise in surface EEG may obscure certain BifMs, and we suggest technical improvements that could enhance detection accuracy. Expanding the dataset and incorporating long-term biological rhythms, such as circadian and multiday cycles, may provide a more comprehensive understanding of seizure dynamics and improve clinical decision-making.