RPD318 - AI-based Clinical Decision Support to Improve End-of-Life Care: Learnings from a Real-World Implementation

Key Message : AI-based tools have the potential to improve end-of-life care by identifying patients at highest risk of mortality in need of advance care planning. Though to be effective, accurate predictions must be paired with systems and resources that engage and enable clinicians to act on these predictions.

Abstract:

Background: AI-based tools have the potential to improve end-of-life care by identifying patients at highest risk of mortality in need of advance care planning (ACP) [1,2]. Yet prediction does not always translate to clinical practice; alert fatigue and insufficient training and time can hinder providers from acting on predictions [3,4].

Objectives: To develop and assess the real-world effectiveness of an AI-based tool that provides holistic support for providers in initiating ACP among high-risk patients.

Methods: We developed an intervention to promote ACP that combines mortality risk estimation using AI, clinician training, person-to-person alert messages, streamlined clinical workflows, and enhanced palliative care capacity. The program was implemented across 8 of 10 adult hospitals in the BJC HealthCare system, and data were collected between December 22, 2020 (at launch in the first hospital sites) and December 31, 2024.

Results: More than 300 clinicians were trained through the program, and 13,976 patients were identified as candidates for ACP. Clinicians exhibited a high response rate (93%) to the patient eligibility alert. Systemwide improvements were observed across several metrics, including
a fivefold increase in ACP notes, from 1.2% of 146,257 encounters in 2021 to 6.7% of 167,681 encounters in 2024, and a 63% increase in the proportion of encounters with palliative care consults, from 2.2% of 146,257 encounters in 2021 to 3.6% of 167,681 encounters in 2024.
The Vizient Mortality Index (a ratio of observed to expected mortality) also decreased by 32% during this time frame (where a lower score indicates that fewer patients died than would have been expected), from 0.92 in 2021 to 0.62 in 2024.
Implications: This case study demonstrates the importance of pairing accurate mortality predictions with systems and resources that enable clinicians to act on these predictions, including engaging with alerts and improving through comprehensive clinical training.

References: [1] Chi S, Guo A, Heard K, Kim S, Foraker R, White P, et al. Development and Structure of an Accurate Machine Learning Algorithm to Predict Inpatient Mortality and Hospice Outcomes in the Coronavirus Disease 2019 Era. Medical Care. 2022;60(5):381-.
[2] Chi S, Kim S, Reuter M, Ponzillo K, Oliver DP, Foraker R, et al. Advanced Care Planning for Hospitalized Patients Following Clinician Notification of Patient Mortality by a Machine Learning Algorithm. JAMA Network Open. 2023;6(4):e238795.
[3] Van der Sijs H, Aarts J, Vulto A, Berg M. Overriding of drug safety alerts in computerized physician order entry. J Am Med Inform Assn. 2006;13(2):138-47.
[4] Fulmer T, Escobedo M, Berman A, Koren MJ, Hernández S, Hult A. Physicians’ views on advance care planning and end-of-life care conversations. J Am Geriatr Soc 2018;66:1201-1205. https://agsjournals.onlinelibrary .wiley .com /doi /10 .1111 /jgs .15374. https://doi .org /10 .1111 /JGS .15374