Multidimensional clustering of clinical, metabolic, and ultrasound features for 30-day mortality risk stratification in ischemic stroke patients with carotid atherosclerotic plaque.

BACKGROUND: Carotid atherosclerotic plaques are a major cause of ischemic stroke, but stenosis-based assessment does not fully capture plaque heterogeneity or prognostic risk. This study aimed to explore data-driven patient phenotypes and preliminary risk stratification by integrating clinical, metabolic, and carotid ultrasound plaque features using unsupervised clustering. METHODS: Ischemic stroke patients admitted within 72 h of onset were retrospectively enrolled from four centers. Baseline demographics, National Institutes of Health Stroke Scale scores, laboratory indicators, carotid ultrasound plaque characteristics (for example, stenosis rate, Plaque Reporting and Data System grade), and 30-day mortality were extracted. After preprocessing and standardization, k-means clustering was applied to baseline features only; 30-day mortality was reserved for post-clustering outcome comparison. The optimal number of clusters was determined using the elbow method and silhouette score, and Uniform Manifold Approximation and Projection was used solely for two-dimensional visualization. Univariable and multivariable logistic regression analyses evaluated whether cluster membership was independently associated with 30-day mortality after adjustment for major clinical, laboratory, and plaque-related confounders. RESULTS: Among 3,764 patients, five exploratory clusters were identified, showing marked heterogeneity in age, neurological severity, metabolic profile, and plaque characteristics. Cluster 0 (29.2%) comprised the youngest patients, with the lowest National Institutes of Health Stroke Scale scores, smallest plaques, and lowest 30-day mortality (2.9%), whereas Cluster 4 (13.6%) had the highest stenosis rates, more unstable plaque features, pronounced metabolic abnormalities, and the highest mortality (11.2%). Overall mortality differences were significant (chi-squared = 12.36, p = 0.015). In multivariable analysis, cluster membership remained independently associated with 30-day mortality: compared with Cluster 0, Cluster 2, Cluster 3, and Cluster 4 showed progressively increased adjusted odds of death. CONCLUSION: Integrating multidimensional clinical, metabolic, and ultrasound features with k-means clustering can identify exploratory patient phenotypes and reveal substantial heterogeneity in short-term risk among ischemic stroke patients with carotid plaques. High-risk clusters with more severe plaque and metabolic abnormalities suggest potential phenotypic heterogeneity in short-term risk, but these findings should be regarded as hypothesis-generating and require external validation before any clinical application can be considered. The retrospective design, reliance on 30-day mortality alone, and limited clustering stability assessment are important limitations; prospective studies with longer follow-up and external validation are needed to confirm clinical utility.