Abstract TMP98: Identifying Stroke Patients At Risk For Atrial Fibrillation Using Electronic Health Record Data And Machine Learning

Abstract

Background: Atrial fibrillation (AF) detection after cryptogenic stroke (CS) or transient ischemic attack (TIA) carries important therapeutic implications. Machine learning (ML) algorithms may help to identify AF risk from electronic health records (EHR). The study aimed to develop and validate a new ML method to identify CS patients at risk for AF using EHR data. Method: We conducted a retrospective cohort study of CS patients from January 1, 2017, to February 28, 2022, utilizing EHR data to query for AF onset based on billing codes. A ML model was established to predict the occurrence of AF during follow-up. The features used in the model included data from 34 variables in EHR linked to AF in other studies. Results: Of 390 CS patients (62±16 years, 56.4% male), 79 (20%) developed AF after index stroke. Of the 34 original variables, ten clinical variables best predicted the occurrence of AF: Age, BMI ≥30, diabetes, dyslipidemia, duration of dyslipidemia, hypertension, chronic kidney disease stage III/IV/V, duration of chronic kidney disease, mitral valve disease, and duration of mitral valve disease. We fitted a cox proportional hazard model to each of the 34 features respectively and ranked them by concordance index. The ten most informative features were selected using a grid search that looped through different combinations of features sequentially and evaluated each combination with 3-fold cross validation. In a Cox proportional hazard model, these ten variables predicted the occurrence of AF with good model discrimination reflected by Harrell’s concordance index of 0.73. Conclusions: ML algorithms using EHR data can be effectively used to identify CS patients at risk for AF. Further study is required to determine if ML algorithms can identify CS patients who may benefit from long-term outpatient cardiac monitoring.