Digital Twin Strata: a Novel Method that Reduces the Sample Size Needed to Analyze Binary Endpoints
A TwinRCT™ includes a robust technique to improve the statistical efficiency of randomized clinical trials. First, historical control arm data is used to train a machine learning model of disease progression. Then, for each subject, the model is applied to their baseline measurements to create a Digital Twin: a longitudinal clinical prediction of their outcomes, in the hypothetical scenario where the subject receives placebo. For continuous outcomes, Digital Twins are incorporated into the trial analysis as a covariate.
This approach reduces the sample size needed to establish efficacy without compromising Type I error control. Other approaches leverage external datasets to improve statistical power, but have an inflated Type I error rate in the presence of unmeasured confounders.
Recent FDA guidance emphasized the need for interpretable, model-free estimands when covariate adjustment is used with non-continuous endpoints. Digital Twin strata are a novel implementation of TwinRCTs with binary endpoints, which provide efficient estimation of marginal risk differences and odds ratios. The Digital Twin strata are used to segment trial subjects, so that we may apply a Cochran–Mantel–Haenszel test.
The sample size reduction from the stratified test depends on the prognostic accuracy of the Digital Twins model. We present simulations that link the reduction to the c-index of the model. Further, we use past clinical trial data to demonstrate the gains that Unlearn.AI’s proprietary TwinRCT can offer in practice.