Prediction of longitudinal outcomes and novel cluster identification in epilepsy

The longitudinal course of epilepsy remains largely unpredictable. This study aimed to predict final outcome and classify dynamic longitudinal trajectories using artificial intelligence. A total of 2586 patients who first visited our epilepsy specialists between 2008 and 2017 and with at least 3 years of follow-up, were retrospectively enrolled. Supervised and unsupervised learning algorithms were employed to identify clusters with distinct longitudinal courses and to examine epilepsy parameters within each cluster. XGBoost showed slightly higher performance than the others for the final outcome prediction. We identified three clusters associated with final seizure freedom and two clusters with persistent seizures. The first cluster demonstrated early remission, often linked to infectious or immune etiologies. Two additional clusters with final seizure freedom exhibited delayed remission. One of these clusters, characterized by relatively lower initial seizure frequency, showed generalized irregular slowing on EEG, and cerebromalacia. The other cluster, with an intermediate seizure frequency, displayed features commonly associated with generalized epilepsy. The fourth cluster displayed ongoing seizures with reduced frequency compared to baseline, with focal spike-and-wave, irregular slowing on EEG, and epilepsy-associated tumor. The fifth cluster experienced consistently high seizure frequency from onset, characterized by hippocampal sclerosis, male predominance, and longer epilepsy duration.