Gastrointestinal bleeding (GIB) is the most common gastrointestinal diagnosis requiring hospitalization in the U.S., and accounts for 2.2 million hospital days and inpatient charges of 19.2 billion dollars. Guidelines recommend using risk stratification for both upper and lower gastrointestinal bleeding to identify very-low-risk patients, which can be defined as patients who do not require a hospital-based intervention. These patients can be considered for discharge from the emergency room with outpatient management, reducing costs without risk of harm to the patient. Existing clinical risk scores for upper and lower GIB are not used in clinical practice due to poor performance and onerous data entry.

We propose a symptom-based deep learning risk score derived from EHR data that bases initial assessment on presenting symptoms, designed for automatic deployment to be available in the relevant time window during the clinical workflow.