University of Pennsylvania, Philadelphia, PA
Manqing Liu , Jinbo Chen , Eric Li , Runze Li , Ravi Bharat Parikh
Background: Machine learning (ML) algorithms outperform traditional tools used for prognostication and may facilitate earlier discussions between oncologists and patients (pts) about hospice enrollment and treatment modification. Identifying longitudinal trajectories of mortality risk may help clinicians and health systems understand which populations such algorithms are likely to benefit. Methods: We identified trajectories of mortality risk and their association with existing metrics of end-of-life care quality, using electronic health and registry data from a prospective cohort of 3,280 pts with cancer who were seen in 18 tertiary or community medical oncology practices within a large academic health system between January 2018 and May 2020 and died prior to November 2020. A validated ML algorithm (c-statistic 0.89; Parikh et al, JAMA Oncol, 2020) prospectively generated mortality risk predictions prior to all encounters. Functional principal component analysis (FPCA) identified modes of variation for all patient-level mortality risk predictions associated with encounters prior to death. Adjusted logistic regression analyses tested associations between mortality risk trajectory and metrics of high-quality end-of-life care. Results: FPCA revealed 2 trajectories that represented 36% and 64% of all pts in the cohort. The first cluster (“unpredictable”) consisted of pts whose ML-predicted mortality risk rose sharply within 30 days of death. The second cluster (“predictable”) consisted of pts whose ML-predicted mortality risk was higher at baseline and rose gradually until death. Individuals with predictable mortality risk trajectories were more likely to have worse performance status, high comorbidity burden, and gastrointestinal (GI) malignancies (Table). Predictable trajectories were associated with higher hospice enrollment (adjusted odds ratio [aOR] 1.87, 95% CI 1.48-2.37), less inpatient death (aOR 0.72, 95% CI 0.56-0.92), less end-of-life intensive care unit admissions in the last 30 days of life (aOR 0.74, 95% CI 0.57-0.95), and less chemotherapy in the last 14 days of life (aOR 0.77, 95% CI 0.55-1.08). Conclusions: Over one-third of deaths among pts with cancer follow an unpredictable trajectory. For pts with unpredictable trajectories, overreliance on ML predictions could perpetuate aggressive end-of-life care.
Characteristics | OR† | P value |
---|---|---|
Age | ||
< 75 (Ref) | ||
≥ 75 | 0.73 | <0.01 |
Comorbidity number | ||
0-1 (Ref) | ||
2 | 1.51 | <0.01 |
3+ | 3.70 | <0.01 |
ECOG performance status | ||
0-1 (Ref) | ||
2+ | 1.29 | 0.03 |
Cancer stage | ||
Stage I-III (Ref) | ||
IV | 2.02 | <0.01 |
Cancer type | ||
Breast (Ref) | ||
GI | 1.56 | 0.03 |
Genitourinary | 1.18 | 0.49 |
Gynecology | 1.37 | 0.43 |
Leukemia | 1.11 | 0.71 |
Lymphoma | 0.63 | 0.06 |
Melanoma | 0.73 | 0.25 |
Myeloma | 0.54 | 0.01 |
Neuro | 0.28 | <0.01 |
Thoracic | 0.87 | 0.50 |
†Odds of belonging to predictable trajectory, relative to unpredictable trajectory.
Disclaimer
This material on this page is ©2024 American Society of Clinical Oncology, all rights reserved. Licensing available upon request. For more information, please contact licensing@asco.org
Abstract Disclosures
2024 ASCO Quality Care Symposium
First Author: Jody S. Garey
2022 ASCO Annual Meeting
First Author: Ravi Bharat Parikh
2023 ASCO Annual Meeting
First Author: Melissa R Rosen
2023 ASCO Quality Care Symposium
First Author: Rushil Virendra Patel