Early Prediction of Delirium in Postcardiac Surgery Patients: Machine Learning Model Development and External Validation

Abstract

Background: Delirium is a frequent postoperative complication among patients who have undergone cardiac surgery and is associated with prolonged hospitalization, cognitive decline, and increased mortality. Early prediction of delirium is therefore critical for initiating timely interventions.

Objective: This study proposes the development and validation of a machine learning-based model to predict postoperative delirium in patients undergoing cardiac surgery during intensive care unit (ICU) care, facilitating the early detection of individuals at high risk of delirium and supporting clinicians in the deployment of targeted preventive strategies.

Methods: This study extracted data on postoperative cardiac surgery patients who remained in the ICU for more than 24 hours from the Medical Information Mart for Intensive Care IV version 2.0 (MIMIC-IV 2.0) database and the eICU Collaborative Research Database (eICU-CRD). The MIMIC-IV 2.0 cohort was randomly divided into a training set and an internal validation set in a 7:3 ratio, whereas the eICU-CRD functioned as an independent validation cohort. We used data from the first 24 hours of ICU monitoring to model the likelihood of delirium over the entire ICU admission period. Delirium was identified by a positive Confusion Assessment Method for the Intensive Care Unit evaluation (ie, score ≥4). We built predictive models by using logistic regression, support vector classifier, extreme gradient boosting (XGB), and random forest classifiers. Their performance was assessed via the area under the receiver operating characteristic curve, accuracy, sensitivity, positive predictive value, negative predictive value, and F₁-score.

Results: The analysis involved 2124 patients from the MIMIC-IV 2.0 database and 2406 from the eICU-CRD. A set of 57variables was selected to construct the predictive models. Among the various machine learning models tested, the XGB model demonstrated the best performance for delirium prediction during internal validation. As for external validation, the model achieved an area under the receiver operating characteristic curve of 0.75, indicating strong discriminatory ability. The most important predictive features identified by the model included hospital length of stay, minimum Glasgow Coma Scale score, mean blood pressure, Sequential Organ Failure Assessment score, weight, urine output, heart rate, and age.

Conclusions: The XGB model with strong predictive capability for ICU delirium after cardiac surgery was developed and externally validated. This model offers essential technical support for building real-time delirium alert systems and enables ongoing risk stratification and evidence-based decision-making within the ICU environment.

Keywords: MIMIC-IV 2.0 database; Medical Information Mart for Intensive Care IV version 2.0 database; cardiac surgery; delirium; eICU Collaborative Research Database; eICU-CRD; intensive care unit; machine learning; prediction model.