Unsupervised clustered-derived phenotyping in carotid artery disease: a novel risk stratification model integrating artificial intelligence algorithm with clinical data

Abstract

Patients with carotid atherosclerotic disease remain at substantial cardiovascular risk despite optimal medical therapy and carotid endarterectomy (CEA). This residual risk may reflect the biological heterogeneity of atherosclerosis. Identifying distinct atherosclerotic phenotypes may improve prediction of Major Adverse Cardiovascular Events (MACE) and enable precision risk stratification. We hypothesize that carotid atherosclerosis comprises distinct clinical phenotypes with differential long-term prognostic implications. We aimed to identify phenotypes integrating clinical variables and artificial intelligence-enabled electrocardiogram (AI-ECG) and to develop a probability score for post-CEA cardiovascular risk stratification. We prospectively included consecutive patients undergoing CEA between 2002 and 2023 with available preoperative AI-ECG and longitudinal clinical data. Patients with >10 missing variables and variables with >30% missing data were excluded. Unsupervised consensus clustering was applied to identify phenotypic classes. Outcomes included first MACE and its components—cardiac events, cerebrovascular events, and all-cause mortality—during long-term follow-up after CEA. The cohort was randomly divided into training and test sets. In the training set, Cox proportional hazards models with clinically guided variable selection were used to derive a probability score discriminating high- versus low-risk phenotypes. Model performance was assessed in the test set by estimating each patient’s probability of high-risk adverse outcomes. The validated model was then applied to the entire cohort, and event-free survival was estimated using Kaplan–Meier analysis. Of 733 patients enrolled, 655 met inclusion criteria (median age 71±9.2 years, 428 [65.3%] males). Two unsupervised phenotypic clusters were identified. In the training set (N=312), one cluster was independently associated with MACE (p=0.026) and all-cause mortality (p<0.001) and was defined as the High-Risk phenotype. A multivariable regression model including demographic, laboratory, and AI-ECG variables predicted the High-Risk phenotype (Figure 1A). A probability threshold of 0.339 assigned patients into High- or Low-risk phenotype. This model demonstrated excellent discrimination (Figure 1B) and was validated in the test set (N=343). Applied to the entire cohort, the model identified patients with significantly lower MACE- and all-cause mortality-free survival over long-term follow-up (Figure 2). The current study demonstrated that, in patients undergoing CEA, the integration of AI-ECG algorithms with clinical data uncovers distinct atherosclerotic phenotypes and identifies a group with elevated long-term cardiovascular risk, supporting a phenotype-driven approach to enhance individualized risk assessment and guide personalized management strategies in carotid atherosclerotic disease.Figure 1 A/BFor image description, please refer to the figure legend and surrounding text.Figure 2For image description, please refer to the figure legend and surrounding text.