The initial outcome of deploying a mortality prediction tool at community oncology practices.

Abstract

1521 Background: Hospice improves the quality of life and care for cancer patients and reduces the likelihood of unwanted death in the hospital. Advance Care Planning (ACP) allows physicians to proactively initiate hospice and end-of-life discussions with identified patients, promoting timely hospice care enrollment. We developed a machine learning (ML) model to predict 90-day mortality risk for patients with metastatic cancer. The tool was designed to enable earlier ACP discussions leading to increased hospice enrollment. This study assesses the ML tool usage on ACP documentation in a community oncology setting. Methods: Twelve practices across the nation were included in the study, all participating in the Oncology Care Model. Five practices implemented the ML tool during 10/26/2020-9/30/2021, with patients scored every two weeks to provide insights on mortality risk. Patients identified as high-risk were evaluated for ACP utilization, obtained from timely EMR data and historical claims. Seven practices did not implement the ML tool and served as the control for the study. Results: A total of 1,663 patients were predicted to have a high risk of mortality at the 12 practices during the timeframe. The median age was 74 years. 53% of patients were males, and 47% were females. ACP documentation varied among the practices. The range was 19.4%-55.8% among ML tool participating practices and 7.4%-31.0% among non-participating practices. The weighted mean of ACP utilization was 34.4% for participating practices and 14.0% for non-participating practices. Compared with non-participating practices, the ACP rate increased significantly by 2.5-fold for participating practices (p-value = 0.03, two-sided T-test). Conclusions: This initial outcome study showed improved ACP documentation after deploying a mortality prediction tool in a community oncology setting. We are currently working on propensity score matching and regression analysis to reduce the effect of confounding factors such as practices, patient demographics, diagnosis, and treatment. Future studies will evaluate the impact of mortality tool use on other outcomes, including hospice enrollment, emergency department visits, and hospital admission. Implementing the mortality prediction tool is an ongoing effort with more practices planned to adopt.