Presentation (CTI)

OBJECTIVES: Pediatric trials often face recruitment constraints that limit sample size and evaluation of alternative endpoint strategies. Synthetic patient generation has been proposed for trial-design simulation, yet its ability to preserve design-relevant statistical properties is rarely validated against ground truth. Using a completed pediatric randomized controlled trial (RCT), we evaluated whether synthetic cohorts reproduce the original trial-design conclusions. We did not aim to establish treatment efficacy or increase evidentiary sample size.
METHODS: Reference data came from a multicenter, double-blind, placebo-controlled rhinopharyngitis trial (254 children, 6-15 years; 2:1 active). Four generators were compared using a pre-specified seven-metric framework assessing distributional fidelity, multivariable structure, treatment-group balance, and real-versus-synthetic distinguishability. The best-performing Gaussian Mixture Model (GMM) was retained. Bootstrap simulations (500 iterations; N=100-2,500) applied the trial ANCOVA (treatment, baseline score, centre, age) to five endpoints: ΔTNSS Day 5, ΔTNSS Day 3, AUC-TNSS, ΔSSCS, and Day-5 cure rate. Effect sizes, confidence intervals, power trajectories, and sample size for 80% power (N80) were estimated using observed and shrinkage-adjusted effects.
RESULTS: The GMM achieved the highest real-versus-synthetic indistinguishability (discriminator AUC=0.31) while preserving endpoint distributions, treatment allocation, and effect-size ranking. ΔTNSS Day 3 showed the largest treatment signal (Cohen's d=0.179; 95% CI −0.08 to +0.44) and lowest N80 (observed-effect N=880; shrinkage-adjusted N≈1,550). The original primary endpoint (ΔTNSS Day 5) showed a smaller, null-consistent effect (d=−0.040) and required N=2,080. Endpoint rankings remained concordant between real and synthetic cohorts. Treatment-group separation peaked early and diminished as symptom trajectories converged.
CONCLUSIONS: Using a trial with known ground truth, the GMM preserved design-relevant properties and reproduced the original trial-design conclusions, supporting synthetic cohorts as a bounded tool for endpoint selection and sample-size planning rather than a substitute for clinical evidence or statistical power. Earlier endpoint Day 3 appeared more statistically efficient than Day 5, requiring prospective confirmation. External validation of synthetic-derived N80 estimates remains necessary.