Multicenter evaluation of prognostic nutritional index and systemic immune-inflammation index in predicting mortality among critically ill cardiovascular and cerebrovascular patients with varied glucose metabolism: a machine learning-based cohort study.

Background: Critically ill patients with cardiovascular and cerebrovascular diseases face high mortality risks, necessitating precise prognostic tools. Current models lack granularity in assessing glucose metabolic subgroups, while isolated use of the Prognostic Nutritional Index (PNI) and Systemic Immune-Inflammation Index (SII) has limitations. This study evaluates their combined predictive value for mortality across glucose metabolic profiles using machine learning. Methods: We conducted a retrospective cohort study of 1,698 patients from the MIMIC-IV database (2008-2019), stratified by glucose metabolic status: normal glucose regulation (NGR), prediabetes (Pre-DM), and diabetes mellitus (DM). Prognostic associations and discrimination performance were evaluated using Cox regression, Kaplan-Meier analysis, and ROC curves. Machine learning models-including logistic regression, decision tree, random forest, XGBoost, and LightGBM-were developed based on Boruta-selected features to predict 28-day and 90-day all-cause mortality. Model performance was assessed using AUC, accuracy, and F1-score. To externally validate the machine learning models, we incorporated an independent cohort of critically ill cardiovascular and cerebrovascular patients (n = 1,194) from two tertiary hospitals in China: The Affiliated Hospital of Qingdao University and Qingdao Municipal Hospital. Results: Higher PNI was associated with reduced mortality, whereas elevated SII predicted higher mortality risk. The combined PNI-SII model outperformed individual indices across glucose subgroups, showing the best performance in Pre-DM patients (AUC = 0.775 for 28-day mortality). PNI's protective effect was attenuated in the DM group, while SII remained consistently predictive. Machine learning models confirmed PNI and SII as top-ranking mortality predictors, particularly in NGR and Pre-DM populations. External validation demonstrated robust generalizability of the models, with comparable AUCs and calibration metrics across the independent Chinese cohort, supporting cross-center applicability. Conclusion: Integration of PNI and SII improves risk stratification and mortality prediction among critically ill patients with cardiovascular and cerebrovascular diseases, especially those with prediabetes. The machine learning models exhibited strong generalizability when externally validated using real-world data from two tertiary hospitals, underscoring their potential for broader clinical application and personalized decision-making.