DOI: 10.1161/circ.148.suppl_1.15639 ISSN: 0009-7322

Abstract 15639: An Explainable Machine Learning-Based Phenomapping Strategy for Adaptive Predictive Enrichment in Randomized Controlled Trials of Preventive Interventions

Evangelos K Oikonomou, Phyllis M Thangaraj, Deepak L Bhatt, Joseph S Ross, Lawrence H Young, Harlan M Krumholz, Marc A Suchard, Rohan Khera
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Introduction: Randomized controlled trials (RCT) are the cornerstone of evidence-based preventive cardiovascular care but are resource intensive. We propose a novel machine learning (ML) strategy of adaptive predictive enrichment through computational trial phenomaps to optimize RCT efficiency.

Methods: In simulated group sequential analyses of two cardiovascular outcomes RCTs of (1) a drug (pioglitazone vs placebo; IRIS), and (2) a management strategy (intensive vs standard blood pressure reduction; SPRINT), we constructed dynamic phenomaps to infer patient profiles benefiting from the intervention vs control during interim analyses. We examined their ability to adaptively enrich subsequent trial enrollment. We compared the final study size and effect estimates between the adaptive and original study designs.

Results: Across 3 interim analyses, our strategy learned dynamic signatures of individualized benefit. By conditioning prospective participant enrollment on dynamic estimates of personalized benefit, our approach enabled a robust reduction in the trial size across simulations (-18% ± 4.7% in IRIS, p=0.008 [ A ] and -27.4% ± 3.4% in SPRINT, p=0.002 [ B ]), also preserving the original average treatment effect (IRIS: hazard ratio of 0.71 ± 0.01 for pioglitazone vs placebo; vs 0.76 in the original trial [ C ]; SPRINT: hazard ratio of 0.72 ± 0.01 for intensive vs standard systolic blood pressure; vs 0.75 in the original trial; all p<0.01 [ D ]). Hierarchical assessment of safety endpoints revealed no safety concerns with adaptive enrichment, whereas post-hoc review of the enriched populations confirmed that no specific demographic groups were excluded from the analysis.

Conclusions: We present an explainable ML algorithm for adaptive predictive enrichment in RCTs leveraging phenome-wide computational measures of personalized benefit. These findings propose a new paradigm to maximize clinical trial efficiency through dynamic data-driven inference.

More from our Archive