A machine learning method for subgroup analysis of randomized controlled trials

Abstract

We developed a machine learning method for subgroup analyses of randomized controlled trials (RCT), and applied it to the results of the SPRINT RCT for treatment of hypertension. To date, the subgroup analyses mostly focused on detecting associations between certain factors and outcome, in the hope that the results will point out biologically (for example, carriers of a certain mutation) or clinically (for example, smokers) distinct subgroups with different outcomes. This seldom worked in the sense of re-launching the intervention for the detected subgroup only and successfully treating it. In contrast, we propose an empirical and general method to develop a predictive multivariate classifier using the RCT outcomes and baseline data. The classifier identifies patients likely to benefit from the intervention, is not limited to a single factor of interest, and is ready for validation in a subsequent pivotal trial. We believe this approach has a better chance of succeeding in identifying the relevant subgroups because of increased accuracy made possible by the use of multiple predictor variables, and opportunity to use advanced machine learning. The method effectiveness is demonstrated by the analysis of the SPRINT trial.