Red cell distribution width-to-albumin ratio as a potential biomarker for short-term mortality risk in critically ill patients with cerebral hemorrhage: a retrospective study with dual-cohort validation

Background The red cell distribution width-to-albumin ratio (RAR) is a composite biomarker integrating inflammatory, nutritional, and stress status; however, its association with short-term prognosis in patients with intracerebral hemorrhage (ICH) remains unclear. Methods This study was a retrospective dual-cohort study. A total of 2,327 ICH patients from the MIMIC-IV database were included as the derivation cohort, and 428 patients from a tertiary hospital were collected as the external validation cohort. The association between RAR and outcomes was analyzed using multivariable Cox regression, with restricted cubic splines employed to examine non-linear relationships. Multiple machine learning algorithms were utilized to screen key prognostic variables, and a logistic regression-based risk prediction model was constructed. Its discriminative ability and stability were validated in both internal and external cohorts. Results After adjusting for multiple confounders, including demographic characteristics, comorbidities, disease severity, and treatment measures, a higher RAR level remained an independent risk factor for 28-day ICU mortality (adjusted HR = 1.17, 95% CI: 1.08–1.27) and 28-day in-hospital all-cause mortality (adjusted HR = 1.14, 95% CI: 1.05–1.23) in ICH patients. Restricted cubic spline analysis further indicated a significant non-linear dose-response relationship between RAR and these outcomes ( P for non-linearity 0.05). In addition, incorporating RAR significantly improved the predictive performance of six traditional critical illness scoring systems, including APACHE II, SOFA, and SAPS II (AUC improvement ranging from 0.016 to 0.188; all DeLong tests P 0.01). Using five machine learning algorithms, we identified seven key variables—age, RAR, INR, total bilirubin, blood urea nitrogen, aspartate aminotransferase, and systolic blood pressure—to construct a short-term mortality risk prediction model for ICH. This model demonstrated robust discriminative ability in the internal training, internal validation, and external validation sets (AUC values of 0.761, 0.723, and 0.723, respectively), outperforming conventional scoring systems. Conclusion RAR is an independent predictor of short-term mortality risk in patients with ICH. The prediction model incorporating RAR exhibits good discriminative ability and cross-cohort stability, offering a practical tool for early identification of high-risk patients and optimization of management strategies. However, this study has certain limitations, including its retrospective design, limited sample size and single-center source for external validation, and lack of neuroimaging data (e.g., hematoma location/volume). Future prospective multi-center studies are needed to further validate its clinical value.