Get AHEAD of Acquired Epilepsy: An Acute Hospitalization Epilepsy Assessment and Detection Tool

Abstract

Preventing acquired epilepsy will require identifying high-risk patients during the latent window between acute brain injury and the first unprovoked seizure, yet clinicians still lack reliable tools to do so. Current prediction approaches rely on narrow clinical information and ignore both the heterogeneity of underlying diagnoses and the high competing risk of death in acutely ill populations. I developed and evaluated a multimodal prognostic framework, AHEAD (Acute Hospitalization Epilepsy risk Assessment and Detection), which integrates demographic and clinical variables, electroencephalogram (EEG) features, and structured neuroimaging features extracted from radiology reports. The analytic cohort comprised 6,315 adults admitted to the Mass General Brigham health system between 2014 and 2024 who underwent continuous EEG within 7 days of admission for an acute neurologic emergency. The primary outcome was acquired epilepsy, defined as ≥1 unprovoked seizure occurring more than 7 days after the index acute neurologic event. I ascertained outcomes using a staged longitudinal phenotyping framework that combined automated electronic health record (EHR) screening with manual chart adjudication, and estimated time-to-event risk with Fine-Gray subdistribution-hazard models that treated death as a competing event rather than as ordinary censoring. Within 2 years, 549 patients (8.7%) developed acquired epilepsy, 1,882 (29.8%) died, 2,259 (35.8%) were censored, and 1,625 (25.7%) remained alive and event-free. In the full cohort, discrimination improved from a baseline concordance index (C-index) of 0.614 to 0.681 with EEG ii features and 0.691 with combined imaging and EEG features. EEG augmentation produced the clearest gains in encephalopathy-, hemorrhage-, and tumor-related presentations; imaging features alone gave smaller and less consistent improvement. Cumulative incidence analyses revealed substantial heterogeneity across diagnosis groups: tumor-associated presentations had the highest 2-year epilepsy incidence, while septic and metabolic presentations had the highest competing mortality. Together, these results show that risk for acquired epilepsy after acute neurologic hospitalization is shaped by strong clinical heterogeneity and heavy competing mortality, and that EEG-derived features add clinically meaningful prognostic information beyond baseline covariates. Subgroup-aware, competing-risk multimodal modeling is a practical foundation for external validation, risk stratification, and prevention-oriented acquired-epilepsy research that can enrich future antiepileptogenesis trials.