Abstract
BACKGROUND: Admission hypothermia remains a frequent and preventable complication in preterm infants and is associated with increased morbidity and mortality. Early risk stratification may enable timely thermal management and targeted preventive strategies. This study aimed to develop and internally validate multivariable machine learning-based models for predicting admission hypothermia in preterm infants. METHODS: We conducted a retrospective cohort study including consecutively admitted preterm infants (<37 weeks' gestation) at a tertiary neonatal referral center in Southwest China (January 2017-January 2025). Admission hypothermia was defined as an axillary temperature <36.5 °C at NICU admission. The dataset was randomly divided into a training cohort (70%) and a validation cohort (30%). Candidate predictors were selected using least absolute shrinkage and selection operator (LASSO) regression. Six models-logistic regression, decision tree, random forest, support vector machine, artificial neural network, and naïve Bayes-were developed. Model performance was evaluated using discrimination (AUC), calibration, Brier score, and classification metrics. Shapley Additive Explanations (SHAP) were applied to enhance interpretability. RESULTS: Among 346 preterm infants, 154 (44.5%) experienced admission hypothermia. LASSO identified 11 predictors, including gestational age, birth weight, ambient temperature, transport time, inborn status, and preheated incubator use. In the validation cohort, AUCs ranged from 0.78 to 0.86, with logistic regression and artificial neural network demonstrating the highest discrimination (AUC = 0.86). Logistic regression showed favorable calibration and interpretability. SHAP analysis identified lower gestational age, lower birth weight, lower ambient temperature, and longer transport time as the strongest contributors to risk. CONCLUSION: Machine learning-based models using routinely available perinatal and environmental variables can effectively predict admission hypothermia in preterm infants. Logistic regression provided robust performance with strong interpretability, supporting its potential integration into early neonatal risk stratification and targeted thermal management strategies.