Development and Validation of a Metastatic Breast Cancer–Specific Prognostic Model Using CancerLinQ Discovery

PURPOSE: Oncologists struggle to know which patients are near end of life to enable timely transitions to supportive care. We developed an electronic health record-based prognostic model to identify patients with metastatic breast cancer (MBC) at high risk of near-term death. METHODS: For model development, we identified patients with MBC in CancerLinQ Discovery, 2000-2020. We randomly selected one encounter per patient, allocating encounters to training (70%) and test (30%) data sets. We evaluated candidate predictors of death within 30 and 90 days using logistic regression, including age, vital signs, laboratory values, performance status, time since last chemotherapy, and tumor phenotype. We varied the alert rate (ie, high-risk proportion) from 5% to 40% and evaluated performance for 30- and 90-day outcomes at each rate. We conducted external validation using an integrated health system database. RESULTS: We identified 9,270 patients with MBC. Significant predictors of mortality were lower sodium, albumin, or pulse oximetry; higher alkaline phosphatase, aspartate transaminase, white blood cell count, creatinine, or pulse rate; decrease in body mass; poor performance status; opioid use; and recent chemotherapy discontinuation. For 90-day mortality, in the internal test set, models had a prediction accuracy of 72%-83% and a positive predictive value of 41%-76% with an AUC of 0.81. External validation revealed an AUC of 0.68 and a Brier score of 0.321 ± 0.027. CONCLUSION: Clinical variables can predict risk of death within 30 and 90 days for patients with MBC. Further validation and optimization studies are required to maximize clinical utility and acceptability before implementation.