Adolescent Metabolic Syndrome with Age- and Sex-Specific Thresholds: A Laboratory-Enhanced, Temporally Validated Machine Learning Model

Background: Adolescent metabolic syndrome (MetS) represents an important early-life cardiometabolic risk factor for type 2 diabetes and cardiovascular disease, yet its identification remains inconsistent because no internationally accepted paediatric diagnostic standard exists. Many studies default to fixed adult anthropometric thresholds, such as waist circumference cut-offs of ≥102 cm (men) and ≥88 cm (women), despite body composition changing substantially and non-linearly across puberty, systematically altering who is classified as MetS-positive in youth.

Objective: To define adolescent MetS using age- and sex-specific percentile thresholds (Johnson/Ford modified paediatric ATP III criteria, with blood pressure further stratified by age, sex, and height) rather than fixed adult cut-offs, and, on this paediatric-appropriate foundation, to develop and temporally validate a laboratory-enhanced machine learning model evaluated on a post-pandemic holdout absent from prior work with formal ablation to quantify the contribution of each predictor domain.

Methods: Seven pooled NHANES survey cycles (2007–2023) yielded 3,567 adolescents aged 12–19 years with complete diagnostic data, the longest observation window applied to this question to date. MetS was classified using age- and sex-specific 90th percentile thresholds for waist circumference and age-, sex-, and height-specific 90th percentile thresholds for blood pressure. A laboratory-enriched predictor panel spanning insulin resistance, lipid metabolism, purine metabolism, and adiposity was evaluated through a two-stage feature selection process (Elastic Net followed by RFECV) and a four-strategy imbalance-handling ablation, yielding a final eight-variable LightGBM classifier with SMOTE-based correction. The model was trained exclusively on 2007-2020 data (n=3,150), with the August 2021-August 2023 cycle (n = 417) held out entirely as a temporally independent test set following the COVID-19-related interruption in NHANES data collection.

Results: The age- and sex-specific percentile definition identified 20.6% more MetS-positive adolescents than an adult-derived waist-circumference definition applied to the same sample (7.88% vs 6.25% prevalence), confirming that percentile-based paediatric criteria capture a materially larger and more physiologically appropriate at-risk population. On the temporal holdout, the resulting model achieved excellent discrimination (ROC-AUC 0.977, 95% CI 0.959–0.991; PR-AUC 0.868, 0.774–0.943) and a negative predictive value of 0.979, correctly excluding MetS in approximately 98% of screened adolescents. SHAP interpretation and systematic ablation converged on a consistent biological signature: TG/HDL-C ratio, BMI, and HOMA-IR jointly drove predictive performance, with the lipid domain alone accounting for the majority of discriminative signal (PR-AUC fell by 0.373 upon removal), while removing the insulin-resistance domain produced only marginal loss, indicating overlapping rather than absent biological signal. Decision curve analysis confirmed net clinical benefit across screening-relevant threshold probabilities (5–50%).

Conclusion: Defining adolescent MetS using age- and sex-specific percentile thresholds, rather than adult cut-offs, relevantly changes who is detected as at risk and provides a more physiologically valid foundation for risk prediction. Built on this definition, a laboratory-enhanced machine learning model validated on a temporally independent August 2021-August 2023 cohort with formal ablation of each predictor domain sustained high, clinically meaningful discriminative accuracy, offering a more rigorous alternative to prior adult-threshold, random-split adolescent MetS models and supporting its potential as a population-level screening tool pending external, prospective validation.