Clustering analysis based on clinically available data for improving risk stratification and outcome prediction of hypertension

Abstract Background Hypertension (HTN) is a heterogeneous condition with several factors contributing to adverse outcomes. Machine learning can harness complex data to unveil outcome predictors and guide prevention. Purpose To use clustering analysis on clinical data to identify HTN risk profiles and their associations with cardiovascular abnormalities and clinical outcomes. Methods A cohort of 14,840 UK Biobank hypertensive participants without pre-existing heart failure (HF) who underwent cardiovascular magnetic resonance (CMR) assessment was examined using K-means clustering on 79 clinical attributes. The clusters’ association with incident HF, mortality, atrial fibrillation (AF), and major adverse cardiovascular events (MACE) were evaluated, alongside imaging characteristics, with mediation analysis probing the role of imaging in risk differentiation. All post-hoc analyses were adjusted for vascular risk factors (VRFs), including diabetes, hypercholesterolaemia, previous myocardial infarction. Results Three distinct clusters differing considerably in lifestyle habits, cardiovascular risk factors, electrocardiographic parameters, and cardiometabolic profiles were identified. Cluster 1 (lowest risk), predominantly females, exhibited the most favourable metabolic profile and outcomes. Cluster 2 (highest risk), predominantly males, had the most adverse metabolic and cardiovascular profiles, and the highest risk of all adverse outcomes, especially AF and HF (hazard ratio (HR) 1.80 [1.45–2.25] and 1.85 [1.39–2.46]; p <0.005). This group was associated with more eccentric LV hypertrophy (LVH), overall poorer left atrial emptying function (LAEF), bi-ventricular systolic function and LV mechanics, and higher systemic vascular resistance (Figure 1). Cluster 3 (intermediate risk) was a more heterogeneous group with the highest burden of diabetes and adiposity and a moderate rise in AF and MACE risk (HR 1.61 [1.12-2.32] and 1.30 [1.07-1.58]; p <0.05). This cluster was associated with concentric LV remodelling and better indices of cardiac mechanics, but greater impairment in LA booster pump function (active LAEF) than Cluster 2 (Figure 1). Cluster 2's effect on MACE was mostly mediated through LV mass increase (proportion mediated, 67%), followed by changes (reducing) in LV global longitudinal strain (33%), and ejection fraction (25.4%). The indirect effect of CMR changes in the association between cluster 3 and MACE was much smaller with the highest proportion of the effects mediated through increased myocardial wall thickness (26%) but not LV mass, and changes in all the components of LA function. Conclusions Clinical data clustering provides insight into distinct HTN risk profiles, each with specific imaging characteristics and clinical trajectories, hinting at different pathophysiological mechanisms. Enhanced phenotyping can facilitate tailored interventions based on the disease subtype and improve clinical outcomes.Figure 1.