| Description |
- Background: Despite improvements, patients with advanced cancer often do not receive goal-concordant care at the end of life (EOL). Challenges with prognostication contribute to delays ACP conversations and result in more aggressive care, gaps in symptom management, and hospice underuse. Early identification of patients at high risk for mortality provides an avenue to implement targeted ACP conversations.
Methods: We used a validated machine learning model that integrates demographics, lab values, vital signs, and medications to predict in-hospital, 30-day, and 6-month mortality risk for patients with a solid malignancy admitted from the emergency department (ED) to a dedicated solid malignancy oncology unit at Duke University Hospital. This unit includes one oncology and one palliative care attending physician, advance practice providers, and nurse coordinators. Providers received an email when a patient was identified as high risk for mortality. A pre-post study design compared ACP documentation and other EOL healthcare utilization before and after the notification intervention. The pre-intervention cohort included patients hospitalized from 1/7/2019 to 10/25/2019 and the post-intervention cohort from 9/19/2020 to 8/31/2021. We excluded patients admitted to the intensive care unit (ICU) in the first 24 hours. We used chi-square or Fisher’s exact tests for categorical variables and Wilcoxon rank sum tests for continuous variables; we stratified comparisons of categorical variables by physician division using Cochran-Mantel-Haenszel tests.
Results: The pre-intervention cohort comprised 105 hospitalizations and 88 unique patients. Mean (SD) age was 64.9 (11.4); 60.0% (n=63) were White, 3.8% (4) Hispanic/Latino, and 65.7% (69) married. The post-intervention cohort comprised 84 hospitalizations and 77 patients. Mean (SD) age was 66.0 (12.2); 52.4% (44) were White, 2.4% (2) Hispanic/Latino, and 60.7% (51) married. Considering index hospitalizations, an ACP note was written for 2.3% (2) of hospitalizations pre-intervention vs. 80.5% (62) post-intervention (p <0.0001). This relationship held even if the physician was in palliative care (4.1% [2] vs. 84.6% [33]) or oncology (0% [0] vs. 76.3% [29]) (p <0.0001). Inpatient length of stay (LOS), hospice referral, code status change, ICU admission rate or LOS, 30-day readmissions, 30-day ED visits, and inpatient and 30-day deaths did not differ significantly between groups.
Conclusions: Identification of hospitalized cancer patients with high mortality risk via machine learning led to a substantial increase in documented ACP conversations but did not impact EOL healthcare utilization. Further integration of the model in clinical practice is ongoing. Our intervention showed promise in changing clinician behavior, and additional work is needed to evaluate downstream impacts.
|
|---|
