Estimation of Cardiometabolic Risk in Turkish Adolescents Using Different Anthropometric Techniques: Development and Temporal Validation of a Machine Learning Model.

BACKGROUND

Although obesity is a recognized cardiometabolic risk determinant, debate persists over the most appropriate anthropometric technique. This study evaluated different anthropometric techniques for predicting cardiometabolic risk in Turkish adolescents and aimed to develop and temporally validate a machine learning-based prediction model, reported in accordance with the TRIPOD statement.

METHODS

Adolescents from the Türkiye Nutrition and Health Survey in 2010 (n = 1357) and 2017 (n = 561) were included. Anthropometric and biochemical parameters were assessed. BMI z-scores, waist-to-hip and waist-to-height ratios, triponderal mass index (TMI), visceral adiposity index (VAI), and lipid accumulation product (LAP) were calculated.

RESULTS

In ROC analysis, VAI showed the highest discrimination (AUC = 0.747, p < 0.001), followed by LAP (AUC = 0.670). In machine learning, XGBoost achieved the highest training discrimination (ROC-AUC = 0.879), whereas logistic regression was the most stable, with minimal overfitting (ΔAUC = 0.008). The logistic regression model was well calibrated (Brier score = 0.107; calibration slope 1.00 development, 1.03 external). External temporal validation showed comparable discrimination across models (AUC = 0.742-0.757), with logistic regression showing the greatest consistency. The model showed a high negative predictive value (NPV = 91.7%) and was therefore better suited to ruling out than ruling in cardiometabolic risk; positive findings warrant confirmatory testing (PPV = 35.4%). A simple logistic-based formula was derived.

CONCLUSIONS

Machine learning approaches, together with the developed model and simplified formula, may be useful tools for estimating cardiometabolic risk in adolescents and could support a stepwise screening strategy, pending recalibration and prospective validation.