4 Minor Traumatic Head Bleeds, Emergency Department Observation Versus Inpatient Admission: Shorter Stays, Less Expensive, No Increased Risk

Abstract

Compliance with the CMS SEP-1 sepsis bundle is a nationally reported metric and will impact hospital reimbursements. Performance overall is poor with some reporting compliance at 33%. While the translational pathway for prognostic models has been well characterized, few machine learning models have been externally validated. This study aims to be the first deep learning model that is integrated into an operational workflow to successfully detect sepsis in real-time. The primary outcome is SEP-1 bundle compliance and secondary outcomes include inpatient mortality, hospital and emergency department (ED) length of stay, as well as process measures such as time from ED arrival identification of "time zero" (as defined by SEP-1) and time to bundle completion.Methods: March, 2016, an interdisciplinary team consisting of clinicians, data scientists and machine learning experts at a large academic medical center embarked on an innovation pilot to develop a novel machine learning model to detect sepsis, named Sepsis Watch (SW). A computable sepsis definition and deep learning model were developed using a curated dataset capturing over 43,000 inpatient admissions between 10/1/2014 -12/31/2015. Ten sepsis definitions were compared and clinicians agreed on the following: >¼ 2 SIRS criteria, blood culture order, and end organ damage. A deep learning model was built to predict which patients would meet that phenotype of sepsis. The model incorporates 121 clinical variables that are continually collected and assessed in real time on every patient that presents to the ED and the first 6 hours of inpatient admission if applicable. Patient status is updated as being "low risk," "medium risk," "high risk," or "septic." If Sepsis Watch predicts a patient to be "High risk" or "septic" that information is communicated to the ED team. Data collected for operational feedback includes total volume of patients who are deemed "High Risk" or "septic", time of day when the model made these predictions, compliance with antibiotic administration, collection of repeat lactate and obtaining cultures all within 3 hours both individually and as a bundle.Results: Sepsis Watch operated in the pilot phase 11/06/18 -05/05/19 and analyzed over 39,000 ED patients for their risk of sepsis. It predicted that 1,318 patients were "high risk." Of those, 948 progressed to "septic" either in the ED or within 6 hours of admission. Bundle compliance within 3 hours was tracked. For those who were "High Risk" but never "septic," 77% had antibiotics administered, 75% had a repeat lactate collected and 72% had cultures collected in 3 hours. For those deemed "septic", 79% had antibiotics administered within 3 hours, 90% had a repeat lactate collected in 3 hours and 88% had cultures collected in 3 hours. Overall bundle compliance was 55.7% for "High Risk" patients and 67.1% for "Septic" patients.Conclusions: Sepsis Watch is the first deep learning model deployed to successfully detect sepsis in real time with operational workflows developed provide support directly to...