Abstract TP223: Predicting Atrial Fibrillation in Embolic Stroke of Unknown Source

Abstract

Background: A considerable portion of patients with embolic stroke of unknown source(ESUS)are later found to have occult atrial fibrillation (AF). Studies have shown that prolonged outpatient cardiac monitoring increases the chances of AF detection and impacts choice of antithrombotic therapy. The Brown ESUS-AF Score, a prediction tool utilizing age and left atrial enlargement has previously shown promise in predicting AF in ESUS patients. In this study we aimed to externally validate the Brown ESUS-AF score and compare its performance against a random forest classifier. Methods: We included all patients with ESUS who underwent an inpatient stroke evaluation followed by outpatient cardiac monitoring between October 9,2015-March 31,2023. To validate the Brown ESUS-AF score, our cohort was divided between those who developed AF during follow-up versus those who did not. To develop the random forest classifier, we included demographics, medical comorbidities, lab records, echocardiographic and electrocardiographic findings. We did 80/20 train-test split to fine tune the model, and we identified features that had the greatest performance contribution. Results: Of 468 patients, 97(20.7%) patients had occult AF. AF patients were older(68.6±13.0 vs 60.1±15.6;p=0.0001), had a higher left atrial (LA) volume index(30.6±12.3 vs 26.2±10.0;p=0.0001) and larger LA diameter(3.8 ± 0.7 vs 3.5 ± 0.7;p=0.002) compared to those who did not develop AF. We found that a Brown ESUS-AF score of≥2 had lower sensitivity in our patient cohort than the original development cohort (42%vs63%). In comparing our random forest classifier to the Brown ESUS-AF score, our model had higher c-statistic (0.71vs0.64). The most important features we identified are age, LAVI, PR and QT Interval, triglyceride level and platelets. Conclusion: Our machine learning model incorporating 52 variables performed better than the Brown ESUS-AF score in predicting AF in our large and diverse ESUS cohort.