Body composition phenotyping of obesity in children aged 6-18 years: multi-strategy clustering and interpretable machine learning.

Abstract

Background: Body composition heterogeneity in childhood obesity is not fully captured by BMI, motivating operational phenotyping using non-invasive measures.

Aim: To identify body composition-based phenotypes of obesity in children aged 6-18 years using complementary clustering approaches and characterise their discriminative structure through interpretable machine learning.

Subjects and methods: We enrolled 78 obese children (6-18 years) from a single-centre outpatient clinic in a cross-sectional design. Thirteen BIA-derived indices were Z-standardised and analysed using parallel unsupervised strategies (PCA-K-means, HCPC, UMAP-DBSCAN). Cluster quality, stability, and agreement were evaluated by mean silhouette, Jaccard bootstrap (100 resamples), and adjusted Rand index (ARI). Separability was assessed with a random forest classifier using five-fold cross-validation (out-of-fold accuracy; macro/micro AUC), with SHAP for interpretability.

Results: Four phenotypes were consistently supported: low-fat/high-muscle, balanced, high-fat/low-muscle, and mixed high-fat/high-muscle. Cross-validated performance was 85.9% accuracy, macro AUC 0.953, and micro AUC 0.965. Structural metrics were silhouette 0.41 and ARI 0.94, with cluster-wise Jaccard 0.789/0.812/0.835/0.858 (range 0.78-0.86). SHAP prioritised fat mass and BMI; BIA-derived basal metabolic indices contributed as device-estimated outputs, while lean mass and skeletal muscle mass showed opposing contributions.

Conclusions: This exploratory, single-centre framework is hypothesis-generating and requires external validation with outcome-linked markers and imaging before any clinical applicability is considered.