Machine Learning-Based Prediction of Delirium and Risk Factor Identification in ICU Patients with Burns: A Retrospective Observational Study.

Abstract

Background: The incidence of delirium in patients with burns receiving treatment in the intensive care unit (ICU) is high, reaching up to 77%, and has been associated with increased mortality rates. Therefore, early identification of patients at high risk of delirium onset is essential for improving treatment strategies.

Objective: This study aimed to create a machine learning model for predicting delirium in patients with burns during their ICU stay using patient data from the first day of ICU admission and to identify predictive factors for ICU delirium in patients with burns.

Methods: This study focused on 82 patients with burns aged ≥18 years who were admitted to the ICU at Mie University Hospital for 24 h or more between January 2015 and June 2023. Seventy variables were measured in patients upon ICU admission and used as explanatory variables in the ICU delirium prediction model. Delirium was assessed using the Intensive Care Delirium Screening Checklist every 8 h after ICU admission. Ten different machine-learning methods were employed to predict ICU delirium. Multiple receiver operating characteristic curves were plotted for various machine learning models, and the area under each curve (AUC) was compared. Additionally, the top 15 risk factors contributing to delirium onset were identified using Shapley Additive exPlanations analysis.

Results: Among the ten machine learning models tested, logistic regression (AUC: 0.906 ± 0.073), support vector machine (AUC: 0.897 ± 0.056), k-nearest neighbors (AUC: 0.894 ± 0.060), neural network (AUC: 0.857 ± 0.058), random forest (AUC: 0.850 ± 0.074), AdaBoost (AUC: 0.832 ± 0.094), gradient boosting machine (AUC: 0.821 ± 0.074), and Naive Bayes (AUC: 0.827 ± 0.095) demonstrated the highest accuracy in predicting ICU delirium. Specifically, 24-h urine output (from ICU admission to 24 h), oxygen saturation, burn area, total bilirubin level, and intubation upon ICU admission were identified as the major risk factors for delirium onset. Additionally, variables, such as the proportion of white blood cell fractions, including monocytes, methemoglobin concentration, and respiratory rate, were identified as important risk factors for ICU delirium.

Conclusions: This study demonstrated the ability of machine learning models trained with vital signs and blood data upon ICU admission to predict delirium in patients with burns during their ICU stay.

Clinicaltrial: