Machine learning-based identification of cardiovascular risk modifiers in patients with very high LDL cholesterol: A cross-sectional study

Le résumé fourni par la source

Individuals with LDL cholesterol (LDL-C) ≥ 190 mg/dL face a markedly increased risk of major adverse cardiovascular events (MACE). Traditional risk calculators such as SCORE2 and the Framingham risk score often underestimate this risk. In this study, we aimed to apply machine learning techniques to identify clinical and laboratory features associated with the presence of MACE among patients with very high LDL-C. In this single-center, cross-sectional study, 246 patients aged 40 to 80 years with at least 1 documented LDL-C ≥ 190 mg/dL between 2014 and 2025 were evaluated. Comprehensive demographic, clinical, anthropometric, lifestyle and laboratory data - including composite scores such as STOP-BANG and fibrosis-4 (FIB-4) - were collected. MACE was defined as acute coronary syndrome, arterial revascularization, peripheral artery disease, ischemic stroke, or transient ischemic attack. Patients were stratified into MACE and non-MACE groups and both conventional statistical analyses and artificial intelligence-based models were employed. Logistic regression, CatBoost, LightGBM, random forest, and XGBoost models were trained (80% training set) and validated (20% test set) to identify variables contributing to the discrimination between patients with and without MACE. CatBoost achieved the highest area under the receiver operating characteristic curve (AUROC) (0.894), followed by logistic regression (0.886), XGBoost (0.860) and LightGBM (0.845). Variables recurrently identified across models in association with MACE status included STOP-BANG score, older age, hypertension, duration of prior statin use, glycated hemoglobin, triglyceride, neutrophil count, aspartate aminotransferase, absence of peripheral pulses, smoking exposure, creatinine, body mass index, apolipoprotein B, alanine aminotransferase and FIB-4 score. Notably, lipid-lowering therapy was suboptimal in this high-risk cohort: although 61% reported prior treatment, only 28% were on active therapy at evaluation, statin discontinuation was frequent and high-intensity regimens were underutilized, including among patients with MACE. Machine learning models identified multiple clinical and laboratory variables associated with prevalent MACE in patients with markedly elevated LDL-C, integrating traditional cardiovascular determinants with potentially underrecognized modifiers. The STOP-BANG score, as a proxy measure for obstructive sleep apnea risk, emerged as one of the most consistently identified variables across models. Prospective, multicenter studies are needed to validate these findings and further explore their potential relevance for cardiovascular risk assessment in this population.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.