Predicting children and adolescents at high risk of poor health‑related quality of life using machine learning methods.

Background: Existing research has identified health‑related quality of life (HRQoL) is influenced by a multitude of factors among children and adolescents. However, there has been relatively limited exploration of the multidimensional predictive factors (individual characteristics, health risk behaviors, and negative life events) that contribute to HRQoL. This study aimed to develop a nomogram to predict the HRQoL in children and adolescents.

Methods: A total of 12,145 children and adolescents were surveyed using stratified cluster sampling method, randomly divided into a training set (n = 8503) and a validation set (n = 3642). Logistic regression, lasso regression, and random forest models were combined to identify the most significant predictors of HRQoL. A nomogram was constructed using multivariate logistic regression. The receiver operating characteristic curve, k-fold cross-validation, decision curve analysis (DCA), and internal validation were used to assess the accuracy, discrimination, and generalization of the nomogram.

Results: Non-suicidal self-injury, academic burnout, parental abuse, stress, bullying victimization, healthy diet, and sleep were found to be significant predictors of HRQoL. The area under the curve (AUC) of the training set was 0.765, whereas that of the validation data was 0.775. The k-fold cross-validation (k = 10) revealed good discrimination in internal validation (mean AUC = 0.771). The nomogram had good clinical use since the DCA covered a large threshold probability: 5%-89% (in the training set) and 4%-81% (in the validation set).

Conclusions: The nomogram prediction model constructed in this study can provide a reference for predicting HRQoL in children and adolescents.