Abstract 13661: An Adaptive Strategy of Predictive Enrichment to Increase Efficiency of Randomized Clinical Trials Using Machine Learning: A Simulation of the IRIS Trial

Introduction: Randomized controlled trials (RCT) represent the cornerstone of evidence-based medicine but often carry high financial costs and lengthy time frames. Hypothesis: We hypothesized that a machine learning (ML) strategy of predictive enrichment can reduce the size of RCTs through adaptive enrollment based on projected response profiles. Methods: We conducted a post-hoc analysis of individual patient data from the Insulin Resistance Intervention after Stroke (IRIS) trial using dynamic computational trial phenomapping. After a study period of 36 months, we performed serial annual interim analyses ( A ). For this, we designed an algorithm that used a phenotypic trial representation using 59 baseline variables to predict personalized effects of pioglitazone on fatal/nonfatal stroke or myocardial infarction (study outcome). Compared with a complete trial enrollment (100%), in an adaptive fashion, we a priori restricted enrollment to two thirds (66.7%) of eligible participants between each interim analysis based on the projected treatment benefit. Since this was defined by their baseline phenotypic profile, treatment randomization was preserved. We employed these patient subsets as comparators and assessed the effect of pioglitazone on the study outcome through Cox regression models. Results: Compared to the unrestricted enrollment of participants ( blue line ), an ML-guided adaptive strategy of predictive enrichment ( green line ) was associated with defining the same treatment effect size with a lower trial size - 3876 vs 2946 participants ( B-C ). Both approaches appeared to meet statistical significance at 6 years after the trial onset ( D-E ). In contrast, a strategy of random patient selection ( orange line ) did not show significant benefit at the same population size. Conclusions: We developed an adaptive strategy for predictive enrichment to increase efficiency of RCTs using machine learning.