Abstract Atrial fibrillation (AF) affects millions of mostly older Americans. Atrial fibrillation contributes to stroke and weighing stroke risk against risk of bleeding from oral anticoagulants (AC) is central to AF management. AC decision-making has become more complex in recent years with the introduction of several target specific oral AC agents. It remains challenging to advise older AF patients about AC since they are frequently at high risk for stroke and complications from AC, particularly bleeding, and data are limited on real-world outcomes of AC among vulnerable populations. Our project, Detecting bleeding Events using Electronic records for Prediction in Atrial Fibrillation (DEEP AF), will leverage novel biomedical natural language processing (NLP) approaches to integrate bleeding-related information from both the structured and unstructured EHR records and statistical and deep learning approaches to predict risk of bleeding on AF patients on AC therapy. DEEP AF will address present knowledge gaps to accurately identify bleeding events from longitudinal electronic health records and to improve bleeding prediction accuracy and focus on modifiable, clinically relevant risk factors, thereby facilitating future large-scale comparative effectiveness studies evaluating AC agents, and helping clinicians develop better approaches to weighing AC risks and benefits.