Abstract Objectives: Real-world data (RWD) from the routine care of patients with cancer captured through EHRs is a valuable resource for research. Understanding the relationship between characteristics and outcomes of patients treated in the real world and those treated in clinical trials is essential to produce evaluable trial-like populations using RWD in oncology for research and regulatory purposes. Methods: This study used: a) RWD from the Flatiron Health EHR-derived, de-identified, longitudinal database (comprising patient-level structured and unstructured data, curated via technology-enabled abstraction selected from approximately 280 US cancer clinics [~800 sites of care]) and b) patient-level data from three completed RCTs (PALOMA-2, MONALEESA-2, and MONARCH-3) including patients with previously untreated hormone receptor positive (HR+), HER2/neu negative (HER2-) mBC, then separately pooled across the trials into two treatment groups, patients who received aromatase inhibitor monotherapy (AI) or a CDK4/6 inhibitor + AI. Key eligibility criteria were similar across the RCTs and were used to select a real world external cohort (rwEC) initiating AI monotherapy on or prior to 11 Nov 2015 (end of MONARCH-3 enrollment period). Patients from the rwEC were matched separately to the control arm and experimental arm patients from the pooled RCT using propensity score method (PSM). The propensity score was estimated by a logistic regression using baseline covariates of age, race, site of disease (visceral, non-visceral), Eastern Cooperative Oncology Group Performance Status (ECOG PS) (0, 1), and metastatic disease (recurrent, new). The matching ratio was 1:1 without replacement with calipers. Covariate balance was measured by the absolute standardized mean difference (ASMD). Due to the high percentage of missing ECOG PS data, matching was repeated 100 times with imputed ECOG PS. The impact of including additional key covariates for propensity matching such as number of disease sites, bone-only disease, and prior endocrine therapy was assessed. Results: There were 1326 patients with HR+, HER2- mBC selected from the EHR-derived database who received first-line AI therapy and 1827 patients randomized in the RCTs (1106 and 721 patients for experimental and control arms, respectively). With 100 matching iterations, 563 rwEC patients on average (range, 547-572) were matched to the RCTs control arm, and 753 rwEC patients on average (range: 741-761) were matched to the RCTs experimental arm. Prior to matching, the ASMD varied widely across all prespecified baseline covariates (4.3 for the rwEC vs. RCTs control arm, 2.6 for the rwEC vs. RCTs experimental arm). After matching was performed, across all baseline covariates used in the PSM, the ASMD was reduced to be under 0.12 for the rwEC vs. RCTs control arm, and under 0.2 for the rwEC vs. RCTs experimental arm in more than 90% of the matching iterations. Analyses looking at the additional baseline covariates to the propensity matching resulted in similar ASMDs. Conclusions: EHR-derived RWD can be used to generate a cohort of patients with similar baseline characteristics to those treated on RCT. The next step in our trial emulation framework is to analyze the comparability of outcomes between these two matched cohorts. Citation Format: Laleh Amiri-Kordestani, Xin Gao, Shrujal Baxi, Erik Bloomquist, Jonathan Bryan, Lynn Howie, Catherine Keane, Paul G. Kluetz, Christy Osgood, Prashni Paliwal, Donna R. Rivera, James Roose, Julie Schneider, Harpreet Singh, Shenghui Tang, Lijun Zhang, Julia A. Beaver. Generating real-world external comparators for randomized clinical trials (RCTs) in metastatic breast cancer (mBC) using electronic health records (EHRs) [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P2-11-05.
Read full abstract