Ma Jessica E, Kilpatrick Kayla W, Davenport Clemontina A, Walter Jonathan, Acker Yvonne, Setji Noppon, Olsen Maren K, Patel Mihir, Gao Michael, Gardner Matthew, Gollon Jamie, Sendak Mark, Balu Suresh, Casarett David
Department of Medicine (J.E.M, J.W., N.S., D.C.), Duke University School of Medicine, Durham, North Carolina, USA; Geriatrics and Extended Care (J.M.), Durham VA Health System, Durham, North Carolina, USA.
Department of Biostatistics and Bioinformatics (K.W.K., C.A.D., M.K.O., M.G.), Duke University, Durham, North Carolina, USA.
J Pain Symptom Manage. 2025 Dec;70(6):602-612. doi: 10.1016/j.jpainsymman.2025.08.013. Epub 2025 Aug 31.
A poor prognosis is an important trigger for advance care planning (ACP) conversations, but clinicians often overestimate prognosis.
To determine whether ACP note documentation increases by notifying inpatient physicians that a patient is at high risk of mortality.
A pragmatic cluster randomized trial at an academic medical center from September 2021 to December 2022 randomized attending physicians on the inpatient medicine team. An email and page notification was sent to physicians randomized to intervention group for admitted patients at high risk of 30-day and 6-month death based on a machine learning model. The notification recommended to have and document an ACP conversation in the electronic health record (EHR). The primary outcome was documentation of an ACP conversation during hospital admission by the randomized physician. The secondary outcome was ACP note documented by any clinician during the hospital admission. Healthcare utilization outcomes included length of stay and discharge to hospice.
Seventy randomized physicians (35 in each group) cared for 314 unique patients (138 control and 176 intervention) at high risk of mortality. Patients of physicians randomized to the intervention group were more likely to have a documented ACP conversation by the randomized physician compared to the control group (34.7% vs. 19.6%; OR 2.04; 95% CI 1.16-3.59). There was no significant change in ACP documentation by any clinician (52.8% intervention vs. 42.8% control group, OR 1.31; 95% CI 0.81-2.13).
Machine learning mortality model notifications can motivate physicians to document ACP conversations during a hospitalization.
