Matching by OS Prognostic Score to Construct External Controls in Lung Cancer Clinical Trials

Area of Science:

• Oncology
• Biostatistics
• Real-World Evidence

Background:

• External controls (eControls) are non-randomized control arms derived from historical data.
• Confounding between experimental and eControl cohorts is a significant challenge due to lack of randomization.
• Prognostic scores are proposed to balance confounding variables, but their performance in constructing oncology eControls is under-analyzed.

Purpose of the Study:

• To evaluate the performance of prognostic scores in constructing reliable eControls for oncology studies.
• To compare the accuracy of ROPRO, a prognostic score, against propensity scores (5Vars, ROPROvars) in estimating overall survival hazard ratios.

Main Methods:

• Constructed eControls using three methods: ROPRO prognostic score, 5-covariate propensity score (5Vars), and 27-covariate propensity score (ROPROvars).
• Utilized an electronic health record-derived de-identified database.
• Compared performance in estimating overall survival (OS) hazard ratio (HR) for 11 advanced non-small cell lung cancer cases.

Main Results:

• ROPRO eControls exhibited lower OS HR error (MAD 0.072) compared to 5Vars (MAD 0.081) and ROPROvars (MAD 0.087).
• OS HR errors were reduced for all methods in phase III studies.
• ROPRO eControl cohorts included more patients on average (6.54%-11.7%) than propensity score methods.

Conclusions:

• Prognostic scores, particularly ROPRO, reliably reproduce controls compared to propensity scores.
• Prognostic scores accommodate numerous variables without significantly increasing propensity score variability, thus preserving matched patient numbers.
• This study validates prognostic scores as a superior method for constructing robust eControls in oncology.