Potential for using external control arms derived from electronic health records to replace control arms from randomized controlled trials

Abstract

Background: The gold standard clinical trial includes a randomized control; however, oncology drug development is increasingly forced to rely on single arm clinical trials. Interpretation of single arm trials can be challenging. Comparison with external controls (EC) may provide context for study findings.

Objectives: Replicate overall survival (OS) results from multiple RCTs in advanced non‐small cell lung cancer (aNSCLC) using EC groups derived from an electronic health record (EHR).

Methods: Design: All Roche aNSCLC RCTs enrolling from 2011 onward were reviewed. Trial selection criteria included sufficient follow‐up, ≥1 US trial site, and relevant biomarkers captured in both the trial and EC group databases. Trial specific eligibility criteria were applied to the EHR data to form the EC cohorts for each trial. Setting: EC patients were identified in the Flatiron Health (FIH) an EHR‐derived longitudinal database which reflects cancer care received and patient outcomes. The FIH lung cohort contains aNSCLC patients diagnosed from 1/1/11. Exposures: Various Outcome: The primary outcome was OS, defined as time from randomization (trial pts) or treatment initiation (EC pts) to death. Patients were censored at the end of trial follow‐up, FIH analysis cutoff date (10/31/17), or last contact date. Statistical Analysis: Propensity score (PS) methods were used to adjust for differences between the trial population and EC pts. The Kaplan‐Meier (KM) method was used to estimate survival. Adjusted Cox models were used to calculate hazard ratios (HR) and 95% confidence intervals (CI) comparing the trial experimental arms with the EC arms. For each study, HRs and KM plots derived using the EHR EC were compared with the RCT results.

Results: Nine aNSCLC trials met the trial selection criteria. Findings from the first 2 trials showed concordance between the adjusted HR comparing the trial experimental arm with the EC arm with the corresponding RCT HR. For example, trial NCT02008227 reported a HR = 0.79. The estimated HR for the corresponding EC analysis was 0.80 (95% CI: 0.65, 0.98). Similarly, the result within trial NCT01903993 produced a HR = 0.72 compared with the EC derived HR = 0.74 (95% CI: 0.54, 1.00). Results for all 9 trials will be presented.

Conclusions: For the trials investigated, HR estimates based on the EC group closely matched those from the RCT. Hence, control arms derived from contemporaneous EHR data may be useful for interpreting findings on OS from single arm trials in aNSCLC.