Abstracts

Rationale: Post-stroke epilepsy (PSE) is a disabling secondary injury of acute ischemic stroke (AIS) and intracerebral hemorrhage (ICH). Some risk factors of PSE have been established, but are limited to ICD-based criteria, structured data availability or manual data extraction, thereby limiting the predictive models that can be applied across patient populations. Future models may be enhanced by leveraging free text information from the electronic health record (EHR). This project uses AEP based on natural language processing methods (NLP) to address whether: (1) EHR notes can be utilized to identify retrospective cohorts of PSE patients, and (2) NLP can extract acute clinical features that are predictive of PSE.

Methods: We identified a retrospective cohort of adult (age ≥18) AIS and ICH patients admitted to Yale New Haven Hospital (2014-2022) who had continuous EEG ≤ 7 days and > 7 days of follow up. PSE was defined as seizure >7 days from stroke onset. We excluded patients with a history of seizures or epileptogenic condition (e.g., acute brain injury, lesion, tumor). Our goal was to optimize a recently developed AEP for extracting epilepsy-related EHR features (Fernandes et al 2023). We evaluated this AEP from stroke onset through 2 years. We first used the output to build a logistic regression (LR) model to distinguish true and false positives from day 8 to 2 years. We then assessed notes at days 0-7, using domain knowledge to select 6 features relating to acute seizures, surgery, and complications to predict PSE risk. We trained a multivariate LR on our AIS cohort using features with a univariate association of <0.1 and validated this model in our ICH cohort. We evaluated performance by calculating area under the receiver operating curve (AUROC).