Validation of PREdiction of DELIRium in ICu patients (PRE-DELIRIC) model for ICU delirium in general ICU and patients with liver disease: a retrospective cohort study

TL;DR

This study validated a model for predicting ICU delirium in over 3,000 patients, finding it moderately accurate overall but less reliable for patients with liver disease.

Contribution

The study is the largest validation of the PRE-DELIRIC model to date and the first to specifically assess its performance in patients with liver disease.

Findings

The PRE-DELIRIC model showed moderate-to-good discrimination (AUROC 0.74) in predicting ICU delirium.

Calibration was poor, with underprediction in moderate-to-high-risk patients and overprediction in very high-risk patients.

In patients with liver disease, the model had similar discrimination (AUROC 0.73) but significant underprediction of delirium risk.

Abstract

Delirium, a neuropsychiatric disorder characterized by disturbances in attention, cognition, and awareness, is a common complication among intensive care unit (ICU) patients. Several predictive models have been developed that aim to identify patients at high risk of delirium. PRE-DELIRIC (PREdiction of DELIRium in ICu) and its recalibrated version, have been externally validated in several studies, but modest sample sizes have meant uncertainty remains, particularly in patient subgroups. Of particular relevance to our population (as a tertiary liver disease centre), performance in patients with liver disease has not been specifically assessed.

This retrospective cohort study evaluated the PRE-DELIRIC model using data from 3312 adult ICU patients at Cambridge University Hospital, between February 2017 and September 2021. Delirium was primarily defined as either a positive Confusion Assessment Method for the ICU (CAM-ICU) or any new administration of antipsychotic medication. Predictive performance was assessed according to discrimination, measured by the area under the receiver operating characteristic (AUROC) and precision-recall curves; and calibration, as quantified by calibration slope and intercept. We also conducted subgroup analyses in patients with liver disease, sedated patients, and across varying opioid dosing.

Delirium occurred in 32.9% of patients. Overall, PRE-DELIRIC demonstrated moderate-to-good discriminative performance (AUROC 0.74; 95% CI 0.72–0.76); but the model significantly underpredicted delirium incidence for those patients predicted to have moderate-to-high delirium risk (PRE-DELIRIC score 0.2–0.6); and overpredicted for those predicted to be at very high risk (PRE-DELIRIC score > 0.6). Among patients with liver disease (41.6% delirium incidence), discrimination was similar to the overall cohort (AUROC 0.73; 95% CI 0.66–0.81), but calibration was poor, with significant under-prediction of delirium. Discrimination was significantly poorer in both sedated patients and patients receiving high opioid dosing.

This is the largest validation study of the PRE-DELIRIC model to date, and the first to specifically consider patients with liver disease. We found moderate-to-good discriminative predictive performance both overall and in liver disease patients, but calibration was only moderate overall, and significantly under-predicted risk in patients with liver disease. Recalibration of the model and further subgroup-specific adjustments may enhance its utility in clinical practice.

The online version contains supplementary material available at 10.1186/s40560-025-00800-3.

Figures