Individualising intensive systolic blood pressure reduction in hypertension using computational trial phenomaps and machine learning: a post-hoc analysis of randomised clinical trials

Abstract

BACKGROUND: The cardiovascular benefits of intensive systolic blood pressure control vary across clinical populations tested in large randomised clinical trials. We aimed to evaluate the application of machine learning to clinical trials of patients without and with type 2 diabetes to define the personalised cardiovascular benefit of intensive control of systolic blood pressure. METHODS: In SPRINT, a trial of intensive (systolic blood pressure <120 mm Hg) versus standard (systolic blood pressure <140 mm Hg) systolic blood pressure control in patients without type 2 diabetes, we defined a phenotypic representation of the study population using 59 baseline variables. We extracted personalised treatment effect estimates for the primary outcome, time-to-first major adverse cardiovascular event (MACE; cardiovascular death, myocardial infarction or acute coronary syndrome, stroke, and acute decompensated heart failure), through iterative Cox regression analyses providing average hazard ratio (HR) estimates weighted for the phenotypic distance of each participant from the index patient of each iteration. Next, we trained an extreme gradient boosting algorithm (known as XGBoost) to predict the personalised effect of intensive systolic blood pressure control using features most consistently linked to increased personalised benefit, before evaluating its performance in the ACCORD BP trial of patients with type 2 diabetes randomly assigned to receive intensive versus standard systolic blood pressure control. We stratified patients based on their predicted treatment effect, and key demographic groups (age, sex, cardiovascular disease, and smoking). We assessed the presence of heterogeneity with an interaction test, and assessed the performance of the algorithm in a simulation analysis of SPRINT in the presence or absence of an artificially introduced heterogeneous treatment effect. FINDINGS: =0·0184). Subgroup analysis based on age (<65 years: HR 0·89 [95% CI 0·71-1·12]; ≥65 years: 0·85 [0·67-1·09]), sex (male: 0·89 [0·72-1·10]; female: 0·85 [0·65-1·10]), established cardiovascular disease (no: 0·89 [0·70-1·14]; yes: 0·84 [0·67-1·06]), or active smoking (no: 0·85 [0·71-1·02]; yes: 1·01 [0·64-1·60]) did not identify groups with heterogeneity of treatment effect. In a simulation analysis of SPRINT, the proposed algorithm detected groups with heterogeneous treatment effects in the presence, but not absence, of simulated subgroup differences. INTERPRETATION: By use of machine learning to define an individual's personalised benefit through phenotypic representations of clinical trials, we created a practical tool for individualising the selection of intensive versus standard systolic blood pressure control in patients without and with type 2 diabetes. FUNDING: National Heart, Lung, and Blood Institute of the US National Institutes of Health.

Ähnliche Arbeiten