Unsupervised machine learning with Cluster Analysis in patients with rheumatoid arthritis: insights in CV risk phenotypes from a multicentre Italian cohort

Rheumatoid arthritis (RA) patients face increased cardiovascular (CV) risk, yet existing 10-year estimation tools often underperform. This study evaluated whether unsupervised machine learning (uML) clustering can identify distinct RA patient phenotypes with varying major cardiovascular events (MACE) potentially improving risk stratification. A cross-sectional, multi-center cohort of RA patients meeting the 2010 ACR/EULAR classification criteria and without prior CV events was enrolled in January 2019 and followed-up for MACE incidence over 5 years. Clinical and demographic data, including traditional CV risk factors and SCORE2 estimates, were collected. Factor Analysis of Mixed Data (FAMD), a generalization of principal component analysis for mixed data types, was applied in Python (ver.3.9) using Prince (ver.0.7.1). Missing data observations were excluded. Bootstrapped eigenvalue distribution determined the number of FAMD components for clustering, followed by Hierarchical Clustering on Principal Components using Ward’s criterion and Euclidean distance. An XGBoost model assessed feature importance, while ANOVA and Chi-square tests evaluated cluster differences. Among 951 RA patients (mean age 61.8 ± 10.5 years; 81.1% female; median disease duration 120 months), 9 MACE occurred (incidence: 1.89/1000 patient-years). Four clusters emerged: Cluster 1 (high HAQ, steroid use, high cholesterol), Cluster 2 (low disease activity, BMI, diabetes, hypertension, SCORE2), Cluster 3 (predominantly male, < 10-year disease duration, csDMARDs; no MACE cases), and Cluster 4 (longer disease duration, hypertension, high triglycerides, lipid-lowering agents, TNFi; highest MACE incidence).uML clustering identified distinct RA phenotypes, complementing SCORE2 for improved CV risk stratification and aiding in recognizing patients at elevated MACE risk.