Reproducibility of overall survival in metastatic colorectal cancer randomized trials using ARCAD-Derived external...

Synthetic control arms (SCAs) derived from historical trial data can complement randomized controlled trials (RCTs), particularly in oncology where feasibility, ethical, and cost constraints may limit conventional trial designs. However, their validity for overall survival (OS) remains uncertain. We evaluated the feasibility and limitations of constructing SCAs from the ARCAD metastatic colorectal cancer (mCRC) database across multiple treatment lines.

Seven landmark RCTs representing first-, second-, and third-line settings were selected. External control arms were constructed from ARCAD individual patient-level data using propensity score matching based on key clinical and biological variables from a validated ARCAD prognostic score, with adjustment for geographic region and time era when needed. A prespecified two-step benchmarking framework assessed: (1) control-arm OS reproducibility and (2) virtual trial comparisons between matched synthetic controls and original RCT experimental arms. Agreement was evaluated using hazard ratios (HRs) and Z-tests.

A total of 28,022 patients met the inclusion criteria. ARCAD-derived SCAs were successfully constructed for all selected RCTs and closely reproduced control-arm OS. After matching, baseline characteristics were well balanced, with minimal differences between included and excluded patients. Virtual trials showed concordant treatment-effect estimates with the original RCTs, with a median absolute HR deviation of 0.05 and no significant differences by Z-tests (P>.05). Survival outcomes were consistently reproduced across treatment lines.

ARCAD-derived SCAs reliably reproduced control-arm OS and benchmark treatment-effect estimates from landmark RCTs in mCRC. Despite limitations related to residual confounding and missing data, this framework supports benchmarking, trial design, and exploratory analyses when conventional control arms are impractical.