Nomogram and machine learning models for predicting the risk of delirium in ICU patients with NSTEMI

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: CatBoost model discrimination

• Primary Outcome: Development of delirium during ICU stay

• Nnt Or Sensitivity: Moderate discrimination with AUC 0.743, outperformed nomogram (AUC 0.61)

• Confidence Interval: AUC 0.743 (95% CI 0.653-0.833)

Clinical Application

Requires implementation of ML infrastructure, relies on readily available clinical variables Relevant to Australian ICUs managing NSTEMI patients, aligns with ANZICS delirium prevention guidelines, could support quality improvement initiatives ICU patients with NSTEMI at risk of delirium

Abstract

Delirium is a frequent and clinically consequential complication among patients admitted to the intensive care unit (ICU). Early risk stratification in patients with non-ST-segment elevation myocardial infarction (NSTEMI) remains challenging. We aimed to develop and validate prediction models for ICU delirium in NSTEMI patients using Boruta-based feature selection and machine learning (ML) approaches, and to compare ML performance with a nomogram model. We retrospectively identified adult ICU patients (≥ 18 years) diagnosed with NSTEMI in the Medical Information Mart for Intensive Care IV database, and included only those with no delirium within the first 24 hours of admission and no prior history of delirium. A total of 392 patients were included (mean age 72.2 ± 12.6 years; 64% male) and categorized into a delirium group (n = 201) and a non-delirium group (n = 191). Clinical characteristics and laboratory indices were extracted. Participants were randomly split into a training cohort (70%) and a validation cohort (30%). The Boruta algorithm was applied in the training cohort for feature selection and identification of relevant predictors. A nomogram was constructed using Boruta-selected variables. In parallel, the same features were incorporated into ML models; model discrimination and clinical utility were evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis. SHapley Additive exPlanations were used to quantify and visualize feature contributions for the best-performing model. Boruta identified 9 key variables: systolic blood pressure, serum calcium, pulse oximetry, serum sodium, creatinine, serum chloride, anion gap, blood urea nitrogen, and body temperature. In the validation cohort, CatBoost achieved the best discrimination among ML models, with an area under the receiver operating characteristic curve of 0.743 (95% CI = 0.653-0.833). The nomogram demonstrated lower discrimination (validation area under the curve = 0.61). SHapley Additive exPlanations analyses indicated that the Boruta-selected variables contributed differentially to delirium prediction, supporting model interpretability. In ICU patients with NSTEMI, a Boruta-informed CatBoost model showed moderate predictive performance and outperformed a nomogram constructed from the same predictors. This interpretable ML approach may facilitate early identification of patients at high risk of delirium and support timely preventive strategies.

References

1. 1.Zuo, D., & Wang, Y. (2026). Nomogram and machine learning models for predicting the risk of delirium in ICU patients with NSTEMI. Medicine, Article 0000000000049062. https://doi.org/10.1097/MD.0000000000049062