Machine learning insights on stress hyperglycemia ratio and mortality in congestive heart failure patients with cerebrovascular disease: an analysis of the MIMIC-IV database

Abstract Background The stress hyperglycemia ratio (SHR), calculated as Admission Blood Glucose (aBG) / (28.7 × glycated hemoglobin A1c (HbA1c) %) − 46.7, integrates acute and chronic glycemic status and predicts adverse cardiovascular events, yet its prognostic value remains unclear among CHF patients complicated with cerebrovascular disease (CVD). This study aimed to explore the association between SHR and all-cause mortality in this specific Intensive Care Unit (ICU) population and provide evidence for targeted glycemic management. Methods We conducted a retrospective cohort study utilizing the MIMIC-IV database. Among 11,196 ICU patients admitted for CHF between 2008 and 2019, 10,214 patients were excluded due to missing HbA1c or admission blood glucose data, resulting in a final analytical cohort of 982 patients. Restricted cubic splines (RCS) and Kaplan-Meier (K-M) curves were employed to illustrate the dose - response association between the SHR and all-cause mortality. Stratified multivariable Cox regression was utilized to estimate hazard ratios, stratified by cerebrovascular disease status. Multiple machine learning models were developed to predict 30-day all-cause mortality, with CatBoost demonstrating the optimal predictive performance. Log-odds-scale SHapley Additive exPlanations (SHAP) values were calculated to quantify the prognostic impact of each variable. Results A total of 982 CHF patients were enrolled, including 320 with concomitant CVD. RCS revealed a linear positive dose-response association between SHR and 30-day and 365-day all-cause mortality among CHF patients with CVD. K-M curves demonstrated that higher SHR quartiles were associated with worse 30-day and 365-day survival among patients with CVD. After full covariate adjustment, patients in the highest SHR quartile showed significantly elevated 30-day mortality risk [HR = 2.79, 95% CI 1.21–6.42, P = 0.016], with a significant linear trend across SHR quartiles (P for trend = 0.004). Consistently, the highest SHR quartile predicted higher 365-day mortality risk [HR = 2.02, 95% CI 1.12–3.66, P = 0.02], accompanied by an overall positive linear trend [P for trend = 0.014]. No independent prognostic effect of SHR was observed in patients without CVD. SHAP analysis confirmed SHR as a key predictor for short-term all-cause mortality, and the optimal CatBoost model achieved moderate discriminative performance (AUC = 0.77). Conclusion Elevated SHR was independently associated with higher 30-day and 365-day all-cause mortality risk solely in critically ill CHF patients with concomitant CVD. The SHR index enables risk stratification and personalized glycemic management for this high-risk ICU cohort.