This exploratory study uses random forest (RF) to predict optimal treatment resulting in longest overall survival (OS) for patients initiating first or second line of therapy (LOT) for HR+/HER2- metastatic breast cancer (mBC) to build understanding of how machine learning may help inform clinical decision-making.
Flatiron Health electronic health records (EHR) were used. Eligible patients were adult females diagnosed with mBC who received ≥1 LOT for mBC between 15Feb2015 and 31Jan2019. Individual regimens were grouped into hierarchy regimen classes with top three included in this analysis (CDK4/6 inhibitor-based therapy, endocrine therapy and chemotherapy). Study cohort was randomly partitioned 1000 times into 80% training and 20% validation subsets. RF survival models were used to predict optimal regimen in each LOT separately based on baseline demographics and clinical characteristics. The gains in OS from patients who received an estimated optimal regimen vs those who did not were examined using Kaplan-Meier method, parametric survival modeling, and Cox proportional hazards regression, adjusted for baseline characteristics imbalance by inverse probability weighting.
The study cohort included 3965 and 2455 patients with first and second LOT, respectively. Less than 50% of patients in the study cohort received optimal regimen classes. RF models suggested greater use of CDK4/6 therapies to maximize OS: increasing from observed 42.2% to estimated optimal 73.9% in first LOT and from observed 40.5% to estimated optimal 66.5% in second LOT. The OS gain, in terms of restricted mean survival time over a 10-year horizon, was 0.63 and 1.09 years, with hazard ratio (95% confidence interval) 0.81 (0.64, 1.04) and 0.62 (0.46, 0.85), in first and second LOT, respectively.
RF was feasible using oncology EHR data, building the evidence to inform how machine learning may provide recommendations for physicians in choosing treatments that improve outcomes for patients with HR+/HER2- mBC.
© 2020 Published by Elsevier Inc.