Randomized controlled trials often require large sample sizes to obtain statistical power. Faster clinical trials can be achieved by augmenting a small study cohort with external datasets of untreated patients. Unfortunately, statistical methods that merge subjects into the trial placebo arm directly tend to inflate the Type I error rate in the presence of unmeasured confounders.
Prognostic Covariate Adjustment (PROCOVA) maintains Type I error control by leveraging a machine learning model, which has been trained on external control data. For each trial subject, the model generates a Digital Twin: a clinical prediction of the subject’s outcome in the case where they are randomized to the control arm. Inference on the average treatment effect is obtained via an ANCOVA, which adjusts for the Digital Twins as a covariate.
PROCOVA decreases the standard error for the treatment effect by a factor of (1-p^2)^0.5, where p is the correlation between Digital Twins and trial outcomes. We present a reanalysis of a Phase II Alzheimer’s Disease trial, where we achieve a similar standard error to the original study with a 23% sample size reduction for the placebo arm.
|