Distinct metabolic phenotypes in adolescents with obesity identified by unsupervised learning: associations with insulin resistance and resting energy expenditure.

BACKGROUND

Pediatric obesity is characterized by substantial metabolic heterogeneity, including adiposity, insulin resistance, inflammation, and energy metabolism. Traditional categorical classifications may inadequately capture this complexity.

Objective

To identify clinically meaningful metabolic phenotypes using unsupervised clustering and to compare the statistical performance and clinical interpretability of two- and three-cluster solutions.

METHODS

In this cross-sectional study, 758 children and adolescents with obesity underwent anthropometric, biochemical, and indirect calorimetry assessments. Fat-free mass was estimated using validated bioimpedance-based equations, and fat-free mass index (FFMI) was standardized before clustering. K-means clustering (k = 2 and k = 3) was performed using FFMI z-score, BMI-SDS, HOMA-IR, triglycerides, and HDL-cholesterol. Cluster validity was assessed using the Silhouette, Elbow, Davies-Bouldin index, Calinski-Harabasz index, and Gap statistic methods. Validation employed variables not included in clustering, such as glucose-insulin dynamics, inflammatory and hepatic markers, blood pressure, respiratory quotient, and resting energy expenditure (REE). Group comparisons were performed using Kruskal-Wallis and Dunn's post-hoc tests with false discovery rate correction.

RESULTS

Both clustering solutions demonstrated significant separation across metabolic domains. The two-cluster model identified metabolically favourable and unfavourable phenotypes differing in adiposity, insulin resistance, inflammatory burden, hepatic markers, blood pressure, and REE. The three-cluster solution revealed a more granular metabolic stratification with an intermediate phenotype characterized by partial metabolic impairment. Validation confirmed robust differences across physiological variables (all FDR-adjusted p < 0.05). Although the two-cluster solution showed slightly superior internal validity, the three-cluster model provided greater clinical granularity and phenotypic resolution.

CONCLUSIONS

Unsupervised clustering identified biologically coherent metabolic phenotypes in pediatric obesity. While the two-cluster solution provided statistically robust stratification, the three-cluster configuration captured a broader spectrum of metabolic heterogeneity and risk profiles, supporting the application of higher-resolution phenotyping approaches in pediatric endocrinology and metabolic risk assessment.