Presentation (CTI)

OBJECTIVES: External control arms (ECAs) utilizing real-world data are increasingly proposed to supplement or substitute for randomized controls in oncology trials with limited populations. This study evaluated the methodological robustness of synthetic external controls across varying data quality using an ovarian cancer trial simulation framework adapted from the Phase 3 ICON7 trial.
METHODS: A Large Language Model-driven framework parameterized using ICON7 trial characteristics, was used to simulate a 500 patient internal trial population with a target progression-free survival Hazard Ratio of 0.81. Three simulated independent real-world datasets of 1000 patients each were generated to represent three data quality scenarios: Scenario 1 (Balanced), Scenario 2 (Sicker and Older Bias), and Scenario 3 (Missing high-risk subpopulations). Propensity score matching was performed using baseline covariates: Age, ECOG performance status, FIGO staging, and surgical debulking status. Post-matching survival distributions were compared using Kaplan-Meier estimators and log-rank testing.
RESULTS: In Scenario 1, matching successfully balanced all covariates, yielding a non-significant log-rank p-value of 0.219, indicating a valid control. In Scenario 2, baseline database flaws forced exclusion of 66 trial patients and caused residual confounding; this generated a false-positive drug efficacy signal with a significant p-value of 0.007. In Scenario 3, structural omission of advanced high-risk patients in the registry caused a severe positivity violation, rendering the matching and resulting p-value of 0.808 mathematically invalid.
CONCLUSIONS: The validity of ECAs is highly sensitive to quality and representativeness of real-world data. When external data sources exhibit baseline bias or missing subpopulations, propensity score matching cannot remedy structural deficits, creating severe risks of false efficacy findings. These findings highlight the importance of evaluating covariate overlap, positivity and cohort representativeness before constructing real-world ECAs. This simulation framework may provide a practical approach for testing ECA methodologies and assessing fitness for purpose of candidate data sources.