Improving clinical trial efficiency using a machine learning-based risk score to enrich study populations.

Abstract

Prognostic enrichment strategies can make trials more efficient, although potentially at the cost of diminishing external validity. Whether using a risk score to identify a population at increased mortality risk could improve trial efficiency is uncertain. We aimed to assess whether Machine learning Assessment of RisK and EaRly mortality in Heart Failure (MARKER-HF), a previously validated risk score, could improve clinical trial efficiency. Mortality rates and association of MARKER-HF with all-cause death by 1 year were evaluated in four community-based heart failure (HF) and five HF clinical trial cohorts. Sample size required to assess effects of an investigational therapy on mortality was calculated assuming varying underlying MARKER-HF risk and proposed treatment effect profiles. Patients from community-based HF cohorts (n=11 297) had higher observed mortality and MARKER-HF scores than did clinical trial patients (n=13 165) with HF with either reduced ejection fraction (HFrEF) or preserved ejection fraction (HFpEF). MARKER-HF score was strongly associated with risk of 1-year mortality both in the community (hazard ratio [HR] 1.48, 95% confidence interval [CI] 1.44-1.52) and clinical trial cohorts with HFrEF (HR 1.41, 95% CI 1.30-1.54), and HFpEF (HR 1.74, 95% CI 1.53-1.98), per 0.1 increase in MARKER-HF. Using MARKER-HF to identify patients for a hypothetical clinical trial assessing mortality reduction with an intervention, enabled a reduction in sample size required to show benefit. Using a reliable predictor of mortality such as MARKER-HF to enrich clinical trial populations provides a potential strategy to improve efficiency by requiring a smaller sample size to demonstrate a clinical benefit.