Machine learning for the prediction of delirium in elderly intensive care unit patients

Purpose This study aims to develop and validate a prediction model for delirium in elderly ICU patients and help clinicians identify high-risk patients at the early stage. Methods Patients admitted to ICU for at least 24 h and using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) in the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database (76,943 ICU stays from 2008 to 2019) were considered. Patients with a positive delirium test in the first 24 h and under 65 years of age were excluded. Two prediction models, machine learning extreme gradient boosting (XGBoost) and logistic regression (LR) model, were developed and validated to predict the onset of delirium. Results Of the 18,760 patients included in the analysis, 3463(18.5%) were delirium positive. A total of 22 significant predictors were selected by LASSO regression. The XGBoost model demonstrated superior performance over the LR model, with the Area Under the Receiver Operating Characteristic (AUC) values of 0.853 (95% confidence interval [CI] 0.846-0.861) and 0.831 (95% CI 0.815-0.847) in the training and testing datasets, respectively. Moreover, the XGBoost model outperformed the LR model in both calibration and clinical utility. The top five predictors associated with the onset of delirium were sequential organ failure assessment (SOFA), infection, minimum platelets, maximum systolic blood pressure (SBP), and maximum temperature. Conclusion The XGBoost model demonstrated good predictive performance for delirium among elderly ICU patients, thus assisting clinicians in identifying high-risk patients at the early stage and implementing targeted interventions to improve outcome.