Integrative interpretation of the central aortic pressure waveform for cardiovascular outcome prediction: a machine learning approach

Background Radial applanation tonometry enables non-invasive assessment of the central aortic pressure waveform. Indices derived from this waveform by pulse wave analysis (PWA) are associated with hypertension-mediated organ damage and incident cardiovascular (CV) events. However, clinical interpretation of PWA measurements is challenging, due to the large number of PWA variables, their strong interconnectivity and their dependence of brachial blood pressure (BP). Unsupervised machine learning may allow comprehensive interpretation of PWA metrics and clarify its value beyond brachial BP for CV risk stratification. Purpose To identify PWA phenogroups from central pressure wave characteristics as derived by radial applanation tonometry and to evaluate their association with major adverse cardiovascular events (MACE) in a community-based cohort. Methods For this prospective cohort study, 1,105 community-dwelling individuals underwent radial arterial tonometry (mean age, 51.4 years; 49.6% women). Radial waveforms were transformed into central aortic waveforms using a validated transfer function and conventional PWA indices were derived. Incident MACE was adjudicated from general practitioner and hospital records. From 17 candidate PWA indices, seven weakly correlated features were retained for clustering, excluding indexes dependent on brachial BP to minimize dominance of variance from brachial BP in the clustering process. Next, Gaussian Mixture modelling (GMM) was applied to subdivide the cohort based on the seven PWA indexes. Given known differences in PWA between men and women, clustering was performed separately for men and women. Using Cox regression, we examined multivariable-adjusted associations between incident MACE and cluster membership. Results In both sexes, GMM identified three distinct PWA profiles (Figure 1). Cluster assignment correlated significantly with age, systolic and diastolic brachial BP, antihypertensive treatment, history of CV disease and total cholesterol (P<0.05 for all). During a mean follow-up of 11.5±4.5 years, 92 men and 63 women experienced at least one MACE. Individuals assigned to Cluster 3 exhibited the highest unadjusted risk for MACE, whereas those in Cluster 1 had the lowest risk (Figure 2). However, after adjustment for age, brachial BP and other clinical covariables, differences in MACE risk between phenogroups were not statistically significant in either sex (Figure 2). Conclusions Integrative phenogrouping of PWA indices by machine learning identified distinct profiles associated with CV risk. However, the phenogrouping did not provide incremental predictive value beyond traditional CV risk factors in a community setting. While potentially informative for vascular phenotyping, integrative PWA analysis does not appear to improve CV risk stratification in the general population.