Predicting End-Of-Life in Hospitalized Patients to Prompt Advance Care Planning

Abstract

End-of-life planning is a core objective of clinical care for patients with life-limiting disease. Interventions—from palliative care to advance care planning (ACP)—can elucidate patient preferences and align treatment accordingly. Communication is key to provide high-quality end-of-life care, but too few patients receive it. There are tools to aid physicians in identifying patients at risk of dying, but few are general and precise enough to highlight those who may benefit from ACP. The objective of this work was to develop predictive models to identify patients at very high-risk of dying within 60 days of admission to the hospital, and communicate that risk to their physicians. Electronic health record data from thousands of patients were used to validate the approach and develop three similar predictive models which use different data to predict death during the first days of a hospitalization. Each model was implemented to label high-risk patients in the electronic health record. A best-practice alert was developed to notify attending physicians, ask if they agree, and prompt ACP where appropriate. In 2019, 340 admissions were identified with high physician agreement (81.8%) and adoption of ACP (72.3%). This observed ACP rate is higher than comparable patients of lesser risk or all who die within 60 days. Both agreement and ACP were associated with worse survival. These data suggest physicians generally agree with model risk estimates, but override the suggestion to defer ACP when patients are at lesser risk. Following model implementation, the majority of identified patients were correctly high-risk and received ACP notes, typically within two days of admission. This work highlights the potential impacts of applied machine learning in healthcare.