Development and validation of a logistic regression screening and classification model for adolescent study-defined elevated HOMA‑IR using machine learning approaches

Insulin resistance (IR) threatens adolescent health, with delayed intervention increasing adult metabolic disorder risk. Existing diagnostic methods are invasive and complex, limiting large-scale or school-based screening. Here, we developed a non-invasive, scalable IR risk screening and classification model to facilitate early identification/intervention. The study involved adolescents aged 12–20 from the NHANES. Univariate regression and LASSO regression were used to identify independent risk factors. Four ML algorithms determined the optimal model. The calibration chart and DCA are utilized to verify the model. The RCS analysis is carried out for the prediction factors, and the optimal decision threshold is established. A total of 1724 individuals were divided into training set and an internal test set comprising 1206 and 518 individuals. 943 individuals were collected for the historical temporal test set. Among 29 body composition parameters from DXA/BIA, 6 variables were finally selected after LASSO regression. Among the four algorithms, logistic regression (LR) model exhibited superior performance. When compared to the other three models, demonstrating relatively high stability and predictive capability, indicating its optimal status as a forecasting tool for adolescent IR risk. Finally, the integration of the four variables was facilitated by the LR model’s variable importance weight, thereby establishing a prediction nomogram platform for the screening and classification of adolescent IR. A machine learning-based adolescent IR screening and classification model provides optimal threshold for Waist-to-height ratio and trunk fat in adolescents across ages/ genders, facilitating clinical decision-making and early screening, and offering theoretical and technical support for non-invasive metabolic disorder intervention.