Using digital twins to reduce sample sizes while maintaining power and statistical accuracy

Abstract

AbstractBackgroundAlzheimer’s Disease (AD) trials are challenging due to their size, length and costs. Bayesian designs can reduce the required sample sizes by using external data sources to construct prior distributions for treatment effects, so that the same effect sizes can be detected with fewer trial participants. We apply a Bayesian approach based on digital twins ‐ longitudinal, patient‐level placebo records with baseline characteristics matched to those of actual subjects in AD trials. These records are generated by a machine learning model of AD progression that we have trained on data from placebo arms of past clinical trials and from observational studies.MethodWe re‐analyzed a Phase II, double‐blind, placebo‐controlled trial of Resveratol in 119 patients with mild‐to‐moderate AD ‐ Mini‐Mental State Examination (MMSE) scores between 14‐26 ‐ using a method we call PROCOVA+. We focused on the Alzheimer's Disease Assessment Scale ‐ Cognitive (ADAS‐Cog11, primary endpoint), the Clinical Dementia Rating ‐ Sum of Boxes (CDR‐SB) and MMSE.PROCOVA+ leverages the digital twins’ outcomes in two ways: as covariate information on the actual trial participants, and to construct prior distributions for the expected placebo outcomes for each endpoint. These prior distributions are centered on the mean outcomes of all digital twins, with dispersions that reflect how accurately digital twins match actual subjects on similar clinical trials. We determined the necessary sample sizes for PROCOVA+ to achieve the same power as the original study, while fixing a 2:1 randomization ratio between the active treatment and placebo arms. We randomly sampled that number of subjects from the original study and computed estimates and credible intervals for each endpoint.ResultThe total sample size across the two arms was reduced by 17%. The estimates of the treatment effects agreed closely with those of the original study for all three endpoints, with credible intervals which were no wider than the original confidence intervals.ConclusionDigital twins present an opportunity to design and analyze AD trials in a way that maintains statistical accuracy, while reducing the required sample sizes and hence the time and cost associated with recruiting and treating extra subjects.