Development and Validation of a Machine Learning Model for Early Prediction of Delirium in Intensive Care Units Using...

Area of Science:

• Critical Care Medicine
• Machine Learning Applications
• Patient Monitoring

Background:

• Delirium in intensive care units (ICUs) presents a significant challenge to patient outcomes and healthcare efficiency.
• Accurate, real-time delirium prediction is crucial for timely intervention and optimized resource allocation in ICUs.

Purpose of the Study:

• To develop a novel machine learning model for predicting delirium in ICU patients.
• The model exclusively utilizes continuous physiological data and routinely available clinical information.

Main Methods:

• Developed and evaluated machine learning algorithms (Random Forest, Extra-Trees, LightGBM) using clinical and continuous physiological data.
• Validated the model using internal, temporal, and external datasets, including prospective Confusion Assessment Method for the ICU (CAM-ICU) evaluations.
• Assessed clinical utility via decision curve analysis and temporal pattern analysis.

Main Results:

• The developed Random Forest model demonstrated robust performance across internal (AUROC: 0.82), temporal (AUROC: 0.73), and external (AUROC: 0.84) validation.
• Reliable delirium determination was confirmed with a Cohen κ coefficient of 0.81.
• Decision curve analysis indicated a positive net benefit, with model scores increasing as delirium onset approached.

Conclusions:

• A machine learning model using routinely measured variables, including physiological waveforms, was successfully developed for ICU delirium prediction.
• The Random Forest model exhibits consistent and effective performance, applicable to a broad range of ICU patients.
• The model's noninvasive nature minimizes additional risk, enhancing its clinical utility for early delirium detection.