Development and Validation of a Routine Electronic Health Record-Based Delirium Prediction Model for Surgical...

Area of Science:

• Geriatric Medicine
• Computational Medicine
• Surgical Outcomes Research

Background:

• Postoperative delirium (POD) is a frequent complication in older adults undergoing major surgery, leading to adverse outcomes.
• Early identification of high-risk patients is crucial for implementing preventative measures.
• Current prediction models often rely on data collected during hospitalization, delaying timely intervention.

Purpose of the Study:

• To develop and externally validate a machine learning model for predicting POD.
• To utilize routine electronic health record (EHR) data available at hospital admission.
• To create a scalable tool for early identification of patients at risk for POD.

Main Methods:

• Trained and compared four machine learning models (logistic regression, random forest, XGBoost, neural network) using 143 EHR features.
• Utilized data from 7167 POD cases and 7167 matched controls across three hospitals (2014-2021).
• Validated models internally and externally, evaluating performance using Area Under the Receiver Operating Characteristic Curve (AUROC).

Main Results:

• Extreme Gradient Boosting (XGB) models demonstrated superior performance compared to other classifiers.
• XGB models trained on 12 months of preadmission data achieved the highest AUROCs.
• The best XGB model achieved a mean AUROC of 0.79 internally, with external validation AUROCs ranging from 0.69 to 0.74.

Conclusions:

• Routine EHR data can be used to develop effective POD prediction models.
• The developed models show good predictive ability but have limited generalizability across different hospital settings.
• These models offer a potential automated screening tool for identifying at-risk patients upon hospital admission.