Risk prediction of delirium in hospitalized patients using machine learning: An implementation and prospective...

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Predictive modeling for patient outcomes

Background:

• Machine learning models show promise in predicting patient outcomes using electronic health records.
• Limited understanding exists regarding the integration of these models into clinical workflows.

Purpose of the Study:

• To implement and evaluate a random forest-based algorithm for identifying hospitalized patients at high risk for delirium.
• To assess the algorithm's performance within a real-world clinical setting.

Main Methods:

• A random forest algorithm was developed to predict delirium risk upon admission and recalculated on the evening of admission.
• Prospective evaluation involved analyzing 5530 predictions over 7 months.
• Internal medicine patient predictions were compared to expert ratings in blinded and non-blinded settings.

Main Results:

• The algorithm achieved 74.1% sensitivity and 82.2% specificity in clinical application.
• Discrimination (AUC=0.86) was comparable to test datasets, but calibration was poor.
• Predictions showed strong correlation with expert-perceived delirium risk (r=0.81 blinded, r=0.62 non-blinded).

Conclusions:

• The implemented machine learning algorithm demonstrated stable performance in predicting delirium, aligning well with expert assessments.
• Further improvements in calibration are necessary for optimal clinical utility.
• Future research should focus on health professional acceptance of such AI tools.