Prediction of Incident Delirium Using a Random Forest classifier

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Delirium is a severe complication in hospitals linked to adverse outcomes.
• Preventing and detecting delirium early is crucial due to its complexity.
• Existing methods for delirium prediction require enhancement for timely intervention.

Purpose of the Study:

• To develop and validate a machine learning model for predicting incident delirium in hospitalized patients.
• To leverage Electronic Health Record (EHR) data for accurate delirium risk assessment.
• To identify key predictive factors for delirium development in an inpatient setting.

Main Methods:

• Utilized a Random Forest algorithm trained on EHR data from 64,038 inpatient visits.
• Defined incident delirium as the first positive Confusion Assessment Method (CAM) screening after 48 hours.
• Employed an 80%/20% data split for training and validation, with under-sampling for class imbalance.

Main Results:

• The predictive model achieved a high accuracy with an ROC AUC of 0.909 (95% CI 0.898 to 0.921).
• The model effectively incorporated demographic data, comorbidities, medications, procedures, and physiological measures.
• Key predisposing and precipitating risk factors were identified as important variables.

Conclusions:

• Machine learning offers a highly accurate approach for predicting hospital-acquired delirium.
• The developed model has the potential for clinical utility in facilitating earlier interventions.
• Early identification of high-risk patients can mitigate the negative impacts of delirium.