Abstract Body: Background: There is a scarcity of US data about the epidemiology of ischemic stroke etiologies. The stroke etiology data field in the Get-with-the-Guidelines (GWTG) Stroke Registry has a 43% missing rate. We sought to develop a scalable machine-learning model to predict ischemic stroke etiology using variables already collected in the GWTG Stroke registry.

Methods: We linked data from two GWTG Stroke registries from YNHH and MGH from 2015-2020 with non-missing, non-cryptogenic stroke etiology inputs adjudicated by the agreement of at least 2 board-certified vascular neurologists’ review of the EHR data from the acute ischemic stroke hospitalization. We trained machine-learning classifiers using XGBoost to predict non-cryptogenic stroke etiology per 4-level TOAST classification using sequentially increasing number of features along with imputation of missing values: (1) Model-1 with age and sex, (2) Model-2 with all GWTG variables excluding ICD codes, and (3) Model-3 with all GWTG variables including ICD codes. Each model underwent Bayesian hyperparameter optimization with 1,000 trials and 5-fold cross-validation. Model performance was evaluated by the following metrics: AUCROC, accuracy, and Cohen’s kappa.

Results: The cohort included 2,724 patients (median age 70; males = 53%). The number of features was 80 in Model 2 and 2,733 in Model 3. In Model-1, the cross-validation mean AUCROC was 0.64, accuracy 0.45, F1 0.30, and kappa 0.063. Model-2's mean AUCROC was 0.78, accuracy 0.57, F1 0.56, and kappa 0.36. Model-3's mean AUCROC was 0.83, accuracy 0.62, F1 0.62, and kappa 0.44. The performance of the three models with respect to diagnosing each etiology is presented in Figure 1.

Conclusion: We developed and validated a proof-of-concept machine-learning model built using variables available in the GWTG Stroke Registry including embedded ICD codes that yield a predicted ischemic stroke etiology in moderate agreement with board-certified vascular neurologists. Upon further improvement, its application to the National GWTG Stroke Registry could allow automatic population of the stroke etiology variable and facilitate etiologic-specific quality evaluation and epidemiologic studies.