Abstract 350: Measuring the Return on Investment of an Artificial Intelligence (AI) Early Warning System

Abstract

Background: The use of AI has been associated with decreased hospital mortality for at risk patients but the financial benefit is unknown. Aims: We sought to measure the change in mortality and cost in six priority clinical conditions following a multi-center phased implementation of eCART (AgileMD), an AI early warning system which utilizes electronic health record (EHR) data to predict clinical deterioration. Methods: eCARTv2 was integrated into the EHR (Oracle Cerner) at four Midwest hospitals in a rolling two-phase implementation. In Phase I, providers viewed scores on a stand-alone web dashboard. In Phase 2, the output was displayed in the EHR, linked to clinical pathways. Phase 0 consisted of a 12 month baseline. Mortality and inflation-adjusted direct variable cost data were compared across the phases in six predetermined diagnosis related group (DRG) conditions: sepsis, heart failure (HF), chronic obstructive pulmonary disease (COPD), stroke, acute myocardial infarction (AMI) and pneumonia. Results: Between 6/1/17 and 2/29/20, 15217 (20.3%) med-surg encounters met inclusion DRG criteria. eCART elevation was most common in sepsis (62.5%), followed by HF (47.4%), pneumonia (43.7%), AMI (39.0%), stroke (35.5%) and COPD (27.1%). Full eCART implementation was associated with significant cost reductions in sepsis, HF, pneumonia, COPD and AMI of $471-795 per encounter and mortality reductions in sepsis and HF (aOR 0.56 and 0.39, respectively). Conclusions: The mortality and cost benefits of eCART appear to be strongest for sepsis and heart failure, where eCART is most likely to be elevated. Savings for sepsis, HF, pneumonia and COPD are primarily driven by labor (ie LOS).