Development and validation of an insulin resistance prediction model in children and adolescents using machine learning algorithms.

Abstract

Background: Insulin resistance (IR) is a precursor to metabolic disorders like type 2 diabetes and hypertension in children and adolescents. Early detection of IR is critical to prevent severe metabolic complications. IR is influenced by factors such as diet, inflammation, and genetics. However, existing studies often focus on limited populations and overlook dietary factors. This study aimed to evaluate the use of machine learning (ML) models for early IR prediction in children and adolescents, emphasizing accuracy.

Methods: We used physical examination data of children and adolescents aged 6-17 years from the China Health and Nutrition Survey (CHNS) database as the training set and collected routine physical examination data from children and adolescents aged 6-17 years admitted to Nanchong Central Hospital and the Nanchong City Jialing District People's Hospital in Sichuan Province from January 2019 to October 2024 for validation. IR was assessed using the Homeostatic Model Assessment for IR (HOMA-IR) score, with a cutoff of >3.0 indicating IR. Potential predictors included demographic details, lifestyle habits, and blood test results. We conducted univariate logistic regression (LR) analysis to select variables with statistical significance and then constructed and compared the back propagation neural network (BPNN), exhaustive Chi-squared automatic interaction detector (E-CHAID), support vector machine (SVM), and LR models.

Results: The training sample included 827 children and adolescents (281 with IR and 546 without IR), while the test sample included 207 participants. The SVM model demonstrated superior predictive accuracy (91.90% in training and 90.34% in test set) compared to the E-CHAID (77.75% in training and 72.95% in test set), BPNN (75.94% in training and 70.05% in test set), and LR models (76.18% in training and 71.01% in test set). Sensitivity, specificity, Youden's index, and area under the curve (AUC) values also favored the SVM model in both training and test samples.

Conclusions: Compared with the E-CHAID, BPNN, and LR models, the SVM model exhibited superior predictive ability for IR in children and adolescents based on physical examination data that include dietary factors. These findings suggest that the SVM model could serve as a valuable tool for early clinical prediction of IR, potentially aiding in the prevention of type 2 diabetes mellitus (T2DM) and associated metabolic complications. Further research is needed to validate these results in larger and more diverse populations.