FROM REAL-WORLD DATA (RWD) TO IN-SILICO DESIGN: AN INTEGRATED APPROACH TO OPTIMISE STUDY DESIGN AND DECISION MAKING
Author(s)
Huan Zhang, PhD, MSc1, Emmanuelle Boutmy, PhD2, Paolo Messina, MSc3, Meritxell Sabido, MD, PhD4.
1Merck Serono Co., Ltd, Beijing, China, an affiliate of Merck KGaA, Darmstadt, Germany, 2Merck Santé S.A.S., Lyon, France, an affiliate of Merck KGaA, Darmstadt, Germany, 3InSilicoTrials, Trieste, Italy, 4Merck Healthcare KGaA, Darmstadt, Germany.
1Merck Serono Co., Ltd, Beijing, China, an affiliate of Merck KGaA, Darmstadt, Germany, 2Merck Santé S.A.S., Lyon, France, an affiliate of Merck KGaA, Darmstadt, Germany, 3InSilicoTrials, Trieste, Italy, 4Merck Healthcare KGaA, Darmstadt, Germany.
OBJECTIVES: To describe an AI enabled in silico framework that generates high fidelity synthetic patient level data from multicentre electronic medical records (EMR) and uses these data to assess whether real world evidence findings are robust to simulated sampling variability and evaluate alternative plausible scenarios to inform trial design decisions.
METHODS: Synthetic (virtual) patient level datasets were generated from models trained on secondary EMR data from four medical centres in two countries, from a source population broader than the inclusion cohort. Each replicate, a plausible alternative sample, underwent the same analysis (including propensity score methods) and was assessed using metrics: (1) the mean difference between synthetic and RWD estimates, (2) the proportion of replicates reproducing the primary non-inferiority conclusion and (3) coverage, whether the RWD estimate fell within the central 90% of the synthetic estimates distribution. Under alternative plausible scenarios, synthetic populations were analysed by varying baseline covariates to quantify the impact of covariate-distribution changes on the primary outcome.
RESULTS: Across 500 synthetic replicates (~2,000 patients each), the synthetic estimate closely matched the RWD estimate on the endpoint’s prespecified absolute scale, with an absolute difference equal to 15% of the minimal clinically important difference (MCID). The RWD estimate fell within the central 90% of the replicate distribution.
Analyses using approximately 1,000,000-patient synthetic cohorts supported plausible scenario evaluation and subgroup oversampling. Across these scenario runs, effect estimates remained directionally consistent with the primary RWD findings, indicating that conclusions were robust to simulated sampling variability and to plausible design choices.
CONCLUSIONS: An AI-based synthetic data approach can provide scalable in silico replications and alternative plausible analyses to inform robustness and study design exploration in multicentre EMR-based target trial emulations. These virtual populations offer the opportunity to stress-test the treatment effect across clinically realistic variations in patient characteristics and optimise development programme designs.
METHODS: Synthetic (virtual) patient level datasets were generated from models trained on secondary EMR data from four medical centres in two countries, from a source population broader than the inclusion cohort. Each replicate, a plausible alternative sample, underwent the same analysis (including propensity score methods) and was assessed using metrics: (1) the mean difference between synthetic and RWD estimates, (2) the proportion of replicates reproducing the primary non-inferiority conclusion and (3) coverage, whether the RWD estimate fell within the central 90% of the synthetic estimates distribution. Under alternative plausible scenarios, synthetic populations were analysed by varying baseline covariates to quantify the impact of covariate-distribution changes on the primary outcome.
RESULTS: Across 500 synthetic replicates (~2,000 patients each), the synthetic estimate closely matched the RWD estimate on the endpoint’s prespecified absolute scale, with an absolute difference equal to 15% of the minimal clinically important difference (MCID). The RWD estimate fell within the central 90% of the replicate distribution.
Analyses using approximately 1,000,000-patient synthetic cohorts supported plausible scenario evaluation and subgroup oversampling. Across these scenario runs, effect estimates remained directionally consistent with the primary RWD findings, indicating that conclusions were robust to simulated sampling variability and to plausible design choices.
CONCLUSIONS: An AI-based synthetic data approach can provide scalable in silico replications and alternative plausible analyses to inform robustness and study design exploration in multicentre EMR-based target trial emulations. These virtual populations offer the opportunity to stress-test the treatment effect across clinically realistic variations in patient characteristics and optimise development programme designs.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR265
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Personalized & Precision Medicine