Synthetic Data Generation vs. Simulated Treatment Comparison: Assessment of the Performance of a Novel Approach for Unanchored Indirect Treatment Comparisons
Author(s)
Samuel Aballea, MSc, PhD1, Igor Chebuniaev, MSc2, Mikolaj Parkitny, MSc3, Mondher Toumi, Prof., M.D., PhD4.
1InovIntell, Paris, France, 2Inovintell, Tbilisi, Georgia, 3Aix-Marseille University, Marseille, France, 4Clever-Access, Kraków, Poland.
1InovIntell, Paris, France, 2Inovintell, Tbilisi, Georgia, 3Aix-Marseille University, Marseille, France, 4Clever-Access, Kraków, Poland.
OBJECTIVES: Standard methods for unanchored indirect treatment comparisons (UAITC) with individual patient data (IPD) for one treatment and aggregate data for comparators include matching adjusted indirect comparison (MAIC) and simulated treatment comparison (STC). STC can extrapolate outcomes for patient profiles not represented in IPD, unlike MAIC, but relies on strong assumptions regarding patient characteristic-outcome relationships. We previously demonstrated that synthetic data generation (SDG) can perform UAITC while reducing variability compared to MAIC. This study aimed to compare SDG performance against STC.
METHODS: We utilized Protocol T data, a randomized controlled trial comparing three treatments in diabetic macular edema. Experiments consisted in estimating the mean gain in visual acuity (VA, ETDRS letter score) with treatment A in a population representative of treatment group B using IPD for A and C and aggregate data for B. To create population imbalances, some patients were removed in IPD: 50% of patients with VA below median (experiment 1) and all first quartile patients (experiment 2). Patient numbers ranged 152-157 per group. STC employed linear regression. For SDG, machine learning combining variational autoencoder and neural ordinary differential equations generated 10,000 virtual patients treated with A but matching group B characteristics. Bootstrap sampling estimated variability.
RESULTS: VA trajectories demonstrated similarity between real and virtual patients treated with A. The expected mean VA change was +15.5 letters. In experiment 1, SDG yielded mean VA change of 15.6 (SE=0.91) versus STC at 14.7 (SE=0.88). In experiment 2, requiring extrapolation for first quartile patients, SDG produced mean change of 15.3 (SE=2.94) compared to STC at 14.5 (SE=1.34).
CONCLUSIONS: SDG enables UAITC without restrictive assumptions about baseline characteristic-outcome relationships, reducing bias compared to STC. While SDG showed higher variability during extrapolation, this may be necessary to avoid STC bias. These findings confirm that SDG represents a promising methodological advancement for UAITC.
METHODS: We utilized Protocol T data, a randomized controlled trial comparing three treatments in diabetic macular edema. Experiments consisted in estimating the mean gain in visual acuity (VA, ETDRS letter score) with treatment A in a population representative of treatment group B using IPD for A and C and aggregate data for B. To create population imbalances, some patients were removed in IPD: 50% of patients with VA below median (experiment 1) and all first quartile patients (experiment 2). Patient numbers ranged 152-157 per group. STC employed linear regression. For SDG, machine learning combining variational autoencoder and neural ordinary differential equations generated 10,000 virtual patients treated with A but matching group B characteristics. Bootstrap sampling estimated variability.
RESULTS: VA trajectories demonstrated similarity between real and virtual patients treated with A. The expected mean VA change was +15.5 letters. In experiment 1, SDG yielded mean VA change of 15.6 (SE=0.91) versus STC at 14.7 (SE=0.88). In experiment 2, requiring extrapolation for first quartile patients, SDG produced mean change of 15.3 (SE=2.94) compared to STC at 14.5 (SE=1.34).
CONCLUSIONS: SDG enables UAITC without restrictive assumptions about baseline characteristic-outcome relationships, reducing bias compared to STC. While SDG showed higher variability during extrapolation, this may be necessary to avoid STC bias. These findings confirm that SDG represents a promising methodological advancement for UAITC.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
PT27
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Sensory System Disorders (Ear, Eye, Dental, Skin)