Association between estimation of pulse wave velocity and all-cause mortality in critically ill patients with ischemic stroke: a retrospective cohort study and predictive model establishment based on machine learning.

Background: Estimated pulse wave velocity (ePWV) has been established as a simple yet effective tool for assessing arterial stiffness and predicting long-term cardiovascular and cerebrovascular mortality. However, the association between ePWV and poor prognosis in critically ill patients with ischemic stroke (IS) remains understudied. This study aimed to investigate the relationship between ePWV and adverse outcomes in critically ill IS patients.

Methods: We conducted a retrospective cohort study using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.0), stratified by ePWV quartiles. Our primary objective was to examine mortality rates at pivotal timeframes: 30 days, 90 days, and 1-year post-observation. Kaplan-Meier (KM) curves complemented these analyses, along with a Cox proportional hazards model, restricted cubic spline curves (RCS), and subgroup analysis, to comprehensively evaluate the association between ePWV and all-cause mortality. To model the mortality risk, four machine learning algorithms were employed, namely Logistic Regression (LR), Random Forest (RF), XGBoost, and Naive Bayes (NB). Model interpretability was improved using Shapley Additive Interpretation (SHAP) analysis, with calibration validating predictive accuracy. We comprehensively compared four machine learning algorithms (LR, RF, XGBoost, NB) against five clinical risk scores.

Results: Our analysis encompassed a cohort of 1,337 patients, with a male preponderance of 51.6%. The 30-day, 90-day, and 1-year mortality rates were 14.1%, 17.9%, and 23.4%, respectively. The RCS analysis revealed a dose-dependent increase in all-cause mortality risk with higher ePWV levels. Critically ill IS patients in the highest ePWV quartile had significantly higher mortality at all time points compared to lower quartiles. Boruta feature selection identified ePWV as a key predictor. The LR model demonstrated superior accuracy in predicting 30-day mortality, while XGBoost outperformed others for 90-day and 1-year mortality predictions.

Conclusion: Elevated levels of the ePWV demonstrate strong prognostic value for both short- and long-term mortality in critically ill IS patients. Machine learning models incorporating ePWV outperformed traditional clinical scores, suggesting potential utility for risk stratification in acute stroke management.