Delirium is a frequent and serious complication among intensive care unit (ICU) patients with sepsis and is associated with poor outcomes. The authors aimed to evaluate whether estimated pulse wave velocity (ePWV), a noninvasive marker of arterial stiffness derived from age and blood pressure, is associated with the risk of delirium in ICU sepsis patients. In addition, they sought to develop and evaluate a predictive model to identify patients at elevated delirium risk.
The investigation used the MIMIC-IV 3.1 database. Eligible subjects were ICU patients meeting the Sepsis-3.0 criteria. The final analytic sample comprised 24,889 patients. Patients were randomly divided into training and validation cohorts in a 7:3 ratio for model development and evaluation. Delirium was defined using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU).
The primary exposure variable was ePWV, estimated from patient age and blood pressure measurements. The primary outcome was the occurrence of delirium during the ICU stay as determined by CAM-ICU assessments. Predictor features for the model were selected using the Boruta algorithm followed by LASSO regression to reduce dimensionality and identify relevant variables for multivariable modeling and machine learning.
Associations between ePWV and delirium were assessed with Kaplan–Meier survival curves and Cox proportional hazards regression. Restricted cubic spline (RCS) analysis evaluated potential nonlinearity in the ePWV–delirium relationship. Subgroup analyses were performed to assess the consistency of associations across clinical subgroups. For prediction, five machine learning models were trained and compared; model interpretability used SHAP (SHapley Additive exPlanations) values to rank feature importance.
In the cohort of 24,889 ICU sepsis patients, the incidence of delirium was 44.4%. Kaplan–Meier analysis showed statistically significant differences in delirium risk across ePWV levels (p < 0.0001). In multivariable Cox regression, higher ePWV was independently associated with increased risk of delirium: hazard ratio per unit increase in ePWV was 1.018 (95% CI 1.008–1.027). When analyzed by quartiles, patients in the highest ePWV quartile had a 19% higher delirium risk compared with the lowest quartile (HR = 1.193).
Restricted cubic spline analysis demonstrated a nonlinear association between ePWV and delirium (RCS p = 0.019), indicating the relationship is not strictly linear across the ePWV range. Subgroup analyses largely supported the primary association across examined subgroups; however, the association was not statistically significant in the subgroup of patients with cerebral infarction.
Five machine learning classifiers were evaluated for delirium risk prediction. The K-Nearest Neighbor classifier achieved the best discrimination in the validation cohort with an area under the receiver operating characteristic curve (AUC) of 0.736. SHAP-based interpretation of the chosen model identified respiratory failure, acute renal failure (ARF) and ACE inhibitor (ACEI) use among the top features contributing to predicted delirium risk.
In this large retrospective cohort from MIMIC-IV, higher estimated pulse wave velocity (ePWV) was independently associated with an increased risk of delirium in ICU patients with sepsis. RCS analysis suggested a nonlinear association, and the relationship held across most subgroups except for patients with cerebral infarction. A machine learning model incorporating selected clinical features demonstrated acceptable discrimination (best validation AUC 0.736 for K-Nearest Neighbor) and identified key contributors to predicted risk via SHAP. The authors propose that ePWV may serve as a useful marker for delirium risk stratification in sepsis and that the predictive model could facilitate earlier recognition of high-risk patients.
Note: The source abstract reports the study design, main statistical results, model performance, and top predictive features. Detailed methods, full variable lists, model hyperparameters, calibration metrics, and external validation details were not reported in the abstract.