Development of a Machine Learning-Based Predictive Model for Postoperative Delirium in Older Adult Intensive Care...

Area of Science:

• Critical Care Medicine
• Geriatric Medicine
• Data Science in Healthcare

Background:

• Delirium is common in geriatric intensive care unit (ICU) patients, increasing complications and negatively impacting prognosis.
• Early identification of high-risk patients is crucial for effective management.

Purpose of the Study:

• To develop and validate a predictive model for postoperative delirium in older ICU patients.
• To enable timely identification of individuals at high risk for delirium.
• To support clinical decision-making for optimized patient care.

Main Methods:

• Utilized data from MIMIC-IV and eICU-CRD databases for training, internal, and external validation.
• Employed Boruta feature selection and compared logistic regression, support vector classifier, random forest, and XGBoost models.
• Evaluated model performance using AUC, Brier scores, and decision curve analysis for prediction windows of 12h, 24h, 48h, and whole stay.

Main Results:

• The XGBoost model demonstrated superior predictive performance, with AUCs ranging from 0.737 to 0.852 across internal and external validation sets.
• Key predictors included first-day delirium assessment, care unit type, minimum GCS score, APS III, acetaminophen, and NSAIDs.
• The model showed accurate calibration and superior net gain compared to other models.

Conclusions:

• A high-performance XGBoost model for predicting delirium in older ICU patients was successfully developed and validated.
• The model enables early prediction of delirium risk at various time points post-admission.
• Facilitates personalized management strategies and care plans for high-risk patients.