OBJECTIVES: Population-adjusted indirect comparisons (PAIC) are often performed using matching, i.e. matching-adjusted indirect comparison (MAIC), in particular when evidence is based on single-arm trials. MAIC suffers from poor precision when effective sample sizes after weighting are small. We tested a synthetic data generation approach in the context of a non-anchored PAIC using clinical trial data in diabetic macular edema (DME).

METHODS: Our synthetic data generator (SDG) architecture combines variational auto-encoders and neural ordinary differential equations. For this test, we selectively sampled patients from 3 arms (designated as A, B and C) of a randomized controlled trial in DME, to introduce a difference in ETDRS score of 5.6 letters at baseline between arms. The SDG was trained on the A and C samples (N=157+152) and the resulting model was used to generate 10,000 synthetic patients treated with A and with baseline characteristics similar to the treatment B sample. The clinical outcome was the mean change in visual acuity, measured by ETDRS letter score, from baseline to 52 weeks. The variability around the mean was estimated by bootstrapping. Results were compared to MAIC, with sandwich estimator of standard error (SE).

RESULTS: The mean change in visual acuity with treatment A before adjustment was 11.56 (SE: 0.86). The adjusted estimate based on SDG was 13.85 (SE: 1.07). The adjusted estimate based on MAIC was 15.16 (SE: 1.36). The effective sample size was estimated at 68% of original sample size.

CONCLUSIONS: The SDG estimate was more precise than the MAIC estimate. Greater precision gains are expected with lower overlap of patient characteristics between datasets. Other benefits of SDG would be possibilities to impute for missing data, to control for modifiable factors post-baseline and to extrapolate. More simulation studies are planned to assess the SDG method.