Predicting Intensive Care Delirium with Machine Learning: Model Development and External Validation.

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Delirium in the intensive care unit (ICU) presents significant patient risks.
• Proactive interventions are crucial for mitigating negative outcomes associated with delirium.
• Accurate forecasting of ICU delirium can guide timely and effective interventions.

Purpose of the Study:

• To predict intensive care unit (ICU) delirium using machine learning algorithms.
• To leverage routinely collected clinical and physiological data from electronic health records (EHRs).
• To develop models for both general prediction and time-sensitive forecasting of delirium onset.

Main Methods:

• Trained and tested two machine learning models on a multicenter EHR database (2014-2015).
• Externally validated models on two single-center databases (2001-2012 and 2008-2019).
• Developed a "24-hour model" and a "dynamic model" predicting delirium up to 12 hours in advance, comparing performance against a reference model.

Main Results:

• The "24-hour model" achieved a mean Area Under the Receiver Operating Characteristic Curve (AUC) of 0.785, outperforming the reference model (AUC 0.730).
• The "dynamic model" demonstrated a higher mean AUC of 0.845 for predicting delirium 12 hours ahead.
• Both models showed strong calibration (Brier Scores around 0.10-0.11) and maintained performance on external validation datasets.

Conclusions:

• Machine learning models effectively predict ICU delirium using standard EHR data.
• The developed models support accurate and dynamic, time-sensitive forecasting of delirium.
• These findings support the integration of predictive analytics for proactive delirium management in ICUs.