275 Background: Oncologists often struggle to know which patients are near end of life to enable a timely transition to supportive care. We developed a breast cancer-specific prognostic tool, using electronic health record data from CancerLinQ Discovery (CLQD), to help identify patients at high risk of near-term death. We created multiple candidate models with varying thresholds for defining high risk that will be considered for future clinical use. Methods: We included patients with breast cancer diagnosed between 1/1/2000 to 6/1/2020 who had at least one encounter with vital signs and evidence of metastatic breast cancer (MBC). All encounters from 1/1/2000 to 7/5/2020 were included. We used multiple imputation (MI) to impute missing numeric variables and treated missing values as a new level for categorical variables. We sampled one encounter per patient and oversampled within 30 days of death, so that the event rate (death within 30 days of encounter) was about 10%. We randomly divided these patients into training (70%) and test datasets (30%). We evaluated candidate predictors of the event using logistic regression with forward variable selection. Candidate predictors included age, vital signs, laboratory values, performance status, pain score, time since chemotherapy, and ER/PR/HER2 receptor status, and change from baseline and change rate of numeric variables. We obtained a single final model by combining resulted logistic regression model from 10 MI training sets. We evaluated this final model on the MI test sets. We varied the alert threshold (i.e., high-risk proportion) from 5% to 40%. Results: We identified 9,270 patients, representing 586,801 encounters. Significant predictors of mortality were: increased age, decreased age at diagnosis, negative change in body mass index, low albumin, high ALP, high AST, high WBC, low sodium, high creatinine, worse performance status, low pulse oximetry, increased age with increased creatinine, high pain score with no opiates, increased pulse rate, unknown/missing PR, opiate use in past 3 months, and prior chemotherapy in past 1 year but not past 30 days. Candidate models had prediction accuracy of 70-89% and positive predictive value of 31-77%. Conclusions: Demographic and clinical variables can be used to predict risk of death within 30 days of a clinical encounter for patients with MBC. Next steps include selection of a preferred model for clinical use, balancing performance characteristics and acceptability, followed by implementation and evaluation of the prognostic tool in the clinic. Candidate models, varying by threshold or percentage of patients assumed to be at high risk, for the outcome of death within 30 days among patients with metastatic breast cancer.[Table: see text]
AGO Algorithms for the Treatment of Breast Cancer: Update 2021
