A 3-Step Procedure to Apply Historical Data and Deep Learning to Reduce Sample Sizes in RCTs Aligned with Regulatory Guidance
Unlearn recently received a draft qualification opinion from the European Medicines Agency (EMA) novel methodologies program for a 3-step procedure called PROCOVA. An essential direction from EMA was to follow each step as described in the PROCOVA Handbook for the Trial Statistician, which details each step. The procedure assumes a prognostic model (e.g., a deep learning model) has been previously developed from historical data to predict patient outcomes under control conditions. Step 1 is to validate this prognostic score for use in a particular planned trial (the Target Trial). Step 2 is to estimate the sample size and plan the Target Trial using PROCOVA for the primary analysis. Step 3, taking place after Target Trial completion, is to estimate the treatment effect using a linear model while adjusting for the prognostic score.
The published correspondence between Unlearn and EMA focused most heavily on expanding the instructions for the application of Step 2, particularly when reducing the randomized controlled trial (RCT) sample size. For continuous outcome measures, the achievable sample size reduction is a function of the correlation between the observed outcomes in the Target Trial and the prognostic score. As this correlation is unknown during the RCT planning phase, assumptions are required to safely estimate this parameter without sacrificing power.
To avoid overly optimistic assumptions, the correlation coefficient must be estimated from data that were not used to develop the prognostic score. Additionally, several cases exist in RCT datasets that may decrease confidence in the strength of correlation between the prognostic model and the validation data set. To account for these potential situations, a deflation parameter is used. Selection of this parameter is detailed in the Handbook according to guidelines established in the EMA draft qualification opinion, supporting the use of PROCOVA for the design and primary analysis of Phase 2 and Phase 3 RCTs.