Machine learning-based prediction of delirium 24 h after pediatric intensive care unit admission in critically ill children: A prospective cohort study

BACKGROUND: Accurately identifying patients at high risk of delirium is vital for timely preventive intervention measures. Approaches for identifying the risk of developing delirium among critically ill children are not well researched. OBJECTIVE: To develop and validate machine learning-based models for predicting delirium among critically ill children 24 h after pediatric intensive care unit (PICU) admission. DESIGN: A prospective cohort study. SETTING: A large academic medical center with a 57-bed PICU in southwestern China from November 2019 to February 2022. PARTICIPANTS: One thousand five hundred and seventy-six critically ill children requiring PICU stay over 24 h. METHODS: Five machine learning algorithms were employed. Delirium was screened by bedside nurses twice a day using the Cornell Assessment of Pediatric Delirium. Twenty-four clinical features from medical and nursing records during hospitalization were used to inform the models. Model performance was assessed according to numerous learning metrics, including the area under the receiver operating characteristic curve (AUC). RESULTS: Of the 1576 enrolled patients, 929 (58.9 %) were boys, and the age ranged from 28 days to 15 years with a median age of 12 months (IQR 3 to 60 months). Among them, 1126 patients were assigned to the training cohort, and 450 were assigned to the validation cohort. The AUCs ranged from 0.763 to 0.805 for the five models, among which the eXtreme Gradient Boosting (XGB) model performed best, achieving an AUC of 0.805 (95 % CI, 0.759-0.851), with 0.798 (95 % CI, 0.758-0.834) accuracy, 0.902 sensitivity, 0.839 positive predictive value, 0.640 F1-score and a Brier score of 0.144. Almost all models showed lower predictive performance in children younger than 24 months than in older children. The logistic regression model also performed well, with an AUC of 0.789 (95 % CI, 0.739, 0.838), just under that of the XGB model, and was subsequently transformed into a nomogram. CONCLUSIONS: Machine learning-based models can be established and potentially help identify critically ill children who are at high risk of delirium 24 h after PICU admission. The nomogram may be a beneficial management tool for delirium for PICU practitioners at present.