Clustering beyond the binary: Stable machine-learning phenotypes of metabolic syndrome

Metabolic syndrome (MetS) is defined by binary thresholds on waist circumference, triglycerides, HDL-C, blood pressure, and fasting glucose, which can obscure risk gradients and within-group heterogeneity. This study evaluates whether unsupervised clustering using physiology-anchored composite indices produces more stable and clinically interpretable cardiometabolic phenotypes than clustering based on traditional MetS components in a full adult population. Specifically, the analysis tests whether such an approach recovers a graded metabolic risk strata extending beyond the binary MetS definition.

NHANES adults (1999-2023) were analysed after excluding users of antihypertensive, lipid-lowering, or glucose-lowering medications. The final analytic cohort comprised 33,597 adults, with a MetS prevalence of 27.5%. Candidate features included traditional MetS components and derived indices. All feature combinations of size 3-8 were screened using k-means clustering (k = 4) with robust scaling. For each combination, clusters were ranked by MetS prevalence. A feature set passed the screen if the union of the two highest-prevalence clusters (Highest and Medium-High) contained ≥75% of MetS+ and ≤15% of MetS- participants. Retained sets were re-fitted across 10 random seeds and evaluated for stability (bootstrap Jaccard with Hungarian alignment), internal validity (silhouette, Davies-Bouldin index), and external alignment (normalised mutual information, purity). The best-performing Derived set and the Traditional benchmark were selected for detailed analysis.

The Derived set concentrated 79.26% of all MetS⁺ in Highest and Medium-High clusters, while including 14.62% of all MetS^-. The Traditional set captured 27.63% of MetS⁺ and 2.51% of MetS^- in the top two clusters. Internal validity favoured the Derived solution (silhouette 0.274 vs. 0.210; Davies-Bouldin 1.152 vs. 1.366), as did stability (min-mean Jaccard 0.963 vs. 0.920). Heatmaps and radar plots showed two distinct intermediate phenotypes in the Derived set, hypertensive-leaning and dyslipidaemic-leaning, whereas the Traditional solution displayed crossovers and weaker gradients. Demographic associations (sex, ethnicity, age) were consistent with known epidemiology.

Composite indices capturing insulin resistance, atherogenic lipids, visceral adiposity, and hepatic fat yielded more reproducible and interpretable strata than traditional components alone, with markedly higher recall of MetS⁺. These findings suggest improved stratification of metabolic dysfunction across a continuum and may support earlier identification of at-risk individuals not yet meeting binary thresholds, pending validation in independent cohorts.