Refining biomarker-based clustering of cardiovascular inflammatory phenotypes in HIV using Recursive Feature Addition: A comparative evaluation approach.

Abstract

People living with HIV remain at elevated risk for a number of non-communicable diseases, including cardiovascular disease (CVD), driven in part by chronic inflammation. While prior studies have identified inflammatory biomarker patterns linked to CVD in people with HIV, it remains unclear which combinations of biomarkers most effectively predict clinical outcomes. We aimed to develop and evaluate a framework for refining biomarker-based clustering approaches to better capture inflammatory patterns associated with a cardiovascular phenotype (CVP) in people with HIV. We developed and evaluated three recursive feature addition (RFA) models to enhance biomarker-driven clustering of people with and without HIV. Using a 24-marker initial panel of biomarkers chosen for their links to clinical CVP in people with HIV, we compared three models for selective inclusion of 31 additional, exploratory biomarkers: (1) a stepwise additive model evaluating biomarkers cumulatively based on biological relevance; (2) a stepwise additive model evaluating biomarkers individually; and (3) a greedy forward-backward selection model. Each model was assessed using principal component analysis (PCA), cluster stability, biological coherence and association with a CVP and 10-year Atherosclerotic Cardiovascular Disease (ASCVD) risk. All three RFA models generated three, biomarker-derived clusters. Post RFA cluster biomarker composition, model stability and clinical associations of these clusters differed across models. The individual additive model (Model 2) produced the most distinct separation of inflammatory profiles, incorporating 11 additional biomarkers, including, GDF-15, IFN-λ2 and Thrombopoietin). In this model, Cluster 3 was characterised by heightened innate and adaptive immune activation, the highest CVP prevalence (11%) and the strongest association with CVP (adjusted odds ratio (aOR) 2.3, 95% CI 1.04-5.09). We demonstrate that an RFA framework using a stepwise, additive model evaluating biomarkers individually to enhance clustering profiles provides optimal unsupervised clustering of exploratory biomarkers to reveal additional associations between inflammatory patterns and CVP in people with and without HIV.