ASSESSING REPRODUCIBILITY OF REAL-WORLD EVIDENCE ANALYSES USING SYNTHETIC DATA: A VALIDATION STUDY IN RARE BRAF V600E METASTATIC NSCLC
Author(s)
Salma Barkaoui, PhD1, Abir Tadmouri, PhD2, Mohammed BENNANI, PhD3, Jerome Vetillard, PhD4, Hadhami Mejbri, Master4.
1Qualees, PARIS, France, 2Pierre Fabre, Paris, France, 3QUALEES, PARIS, France, 4Qualees, Paris, France.
1Qualees, PARIS, France, 2Pierre Fabre, Paris, France, 3QUALEES, PARIS, France, 4Qualees, Paris, France.
OBJECTIVES: Rare oncology populations often provide limited sample sizes for robust real-world evidence (RWE), health economics and outcomes research (HEOR), and health technology assessment (HTA). This study evaluated whether identical analyses applied to real and synthetic BRAF V600E metastatic non-small cell lung cancer (mNSCLC) cohorts yield concordant survival, prognostic, and subgroup conclusions.
METHODS: Source data comprised 184 patients from OCTOPUS, a multinational real-world study across five European countries before approval of encorafenib plus binimetinib. Gaussian Copula (Ledoit-Wolf covariance shrinkage for n<500), CTGAN, and TVAE were compared. A synthetic population of 10,000 patients generated N=184 comparison cohorts. Validation included distributional fidelity (Kolmogorov-Smirnov, Wasserstein distance, Total Variation Distance; bootstrap 95% CI, N=200), machine-learning utility, Kaplan-Meier with log-rank and 12-month landmark analyses, Cox multivariable analyses (ECOG performance status, sex, brain metastases, age, smoking status, treatment lines) across 100 synthetic resamples (N=184 each), and subgroup analyses.
RESULTS: Gaussian Copula achieved the highest fidelity (0.954; bootstrap 95% CI: 0.941-0.957) and ML utility (99.7%), outperforming CTGAN (0.802) and TVAE (0.645). Baseline demographic and clinical characteristics were well preserved, with comparable distributions for sex, histology, brain metastasis status, PD-L1, and smoking history. Median OS was 33.4 months versus 31.7 months and median rwPFS 8.5 versus 8.3 months (synthetic vs. real). Absolute 12-month OS difference was +8.9%, with no significant log-rank differences (p>0.05). Survival concordance was maintained across ECOG performance status, sex, and brain metastasis subgroups. Cox analyses preserved the direction and relative importance of all prognostic factors (ECOG ≥2: HR 2.75 vs. 1.21; brain metastases: 1.99 vs. 1.59; male sex: 1.75 vs. 1.52). No prognostic factor reversed direction.
CONCLUSIONS: Identical analyses of real and synthetic data yielded concordant survival, subgroup, and prognostic conclusions. Gaussian Copula reproduced key characteristics of a rare 184-patient BRAF V600E mNSCLC cohort, supporting synthetic data for evidence augmentation, privacy-preserving data sharing, and HEOR/HTA in rare oncology
METHODS: Source data comprised 184 patients from OCTOPUS, a multinational real-world study across five European countries before approval of encorafenib plus binimetinib. Gaussian Copula (Ledoit-Wolf covariance shrinkage for n<500), CTGAN, and TVAE were compared. A synthetic population of 10,000 patients generated N=184 comparison cohorts. Validation included distributional fidelity (Kolmogorov-Smirnov, Wasserstein distance, Total Variation Distance; bootstrap 95% CI, N=200), machine-learning utility, Kaplan-Meier with log-rank and 12-month landmark analyses, Cox multivariable analyses (ECOG performance status, sex, brain metastases, age, smoking status, treatment lines) across 100 synthetic resamples (N=184 each), and subgroup analyses.
RESULTS: Gaussian Copula achieved the highest fidelity (0.954; bootstrap 95% CI: 0.941-0.957) and ML utility (99.7%), outperforming CTGAN (0.802) and TVAE (0.645). Baseline demographic and clinical characteristics were well preserved, with comparable distributions for sex, histology, brain metastasis status, PD-L1, and smoking history. Median OS was 33.4 months versus 31.7 months and median rwPFS 8.5 versus 8.3 months (synthetic vs. real). Absolute 12-month OS difference was +8.9%, with no significant log-rank differences (p>0.05). Survival concordance was maintained across ECOG performance status, sex, and brain metastasis subgroups. Cox analyses preserved the direction and relative importance of all prognostic factors (ECOG ≥2: HR 2.75 vs. 1.21; brain metastases: 1.99 vs. 1.59; male sex: 1.75 vs. 1.52). No prognostic factor reversed direction.
CONCLUSIONS: Identical analyses of real and synthetic data yielded concordant survival, subgroup, and prognostic conclusions. Gaussian Copula reproduced key characteristics of a rare 184-patient BRAF V600E mNSCLC cohort, supporting synthetic data for evidence augmentation, privacy-preserving data sharing, and HEOR/HTA in rare oncology
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD99
Topic
Health Technology Assessment, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Reproducibility & Replicability
Disease
Oncology