Explainable machine learning for identifying insufficient sleep duration among preschool children in cold-region China: a cross-sectional study

Preschool sleep health is related to movement behaviors, body composition, and motor development. Cold-region environments can alter outdoor opportunities and the seasonal organization of movement behaviors, but current evidence does not establish a uniform cold-region sleep phenotype. We developed and internally validated explainable models to examine whether routinely collected behavior, body-size, and fitness indicators jointly characterize caregiver-reported insufficient sleep duration among preschool children in Heilongjiang, China. This cross-sectional prediction-model study analyzed 2020 monitoring and caregiver-questionnaire data from 1,258 children aged 3–6 years. After excluding variables with > 15% missingness, participants underwent an outcome-stratified 7:3 split. A multivariate imputation by chained equations (MICE) model was fitted in the training set and applied to the internal validation set without refitting. Standardization, least absolute shrinkage and selection operator (LASSO) feature selection, and hyperparameter tuning were restricted to training data. Seven classifiers were evaluated using discrimination, classification metrics, calibration, and decision curve analysis. The selected random forest (RF) model was interpreted using Shapley additive explanations (SHAP). Among 1,258 children, 344 (27.3%) had insufficient sleep duration. LASSO retained body mass index (BMI), weekend moderate-to-vigorous physical activity, weekend outdoor physical activity, weekend screen time, standing long jump, and two-leg consecutive jumping time. Light Gradient Boosting Machine had the highest area under the receiver operating characteristic curve (AUC, 0.849; 95% confidence interval [CI], 0.806–0.889). RF showed numerically higher discrimination than logistic regression (AUC, 0.827 [95% CI, 0.781–0.872] vs. 0.752 [95% CI, 0.692–0.806]) and a balanced profile (accuracy, 0.798; sensitivity, 0.553; specificity, 0.891; F1 score, 0.600; Brier score, 0.143). SHAP ranked weekend outdoor activity, weekend moderate-to-vigorous physical activity, and BMI highest and showed nonlinear feature-output patterns. Routine monitoring data supported an interpretable nonlinear multivariable framework for integrating behavior, body size, and physical fitness in the characterization of caregiver-reported insufficient sleep duration. Its main value is multidimensional data integration, status characterization, and hypothesis generation for prospective studies rather than replacement of direct sleep assessment. External validation across regions, climates, and seasons is required.