Predicting 28-day all-cause mortality in critically ill ischemic stroke patients using white blood cell-to-hemoglobin ratio and dual-feature selection machine learning models

Background The white blood cell-to-hemoglobin ratio (WHR) is a composite biomarker of inflammation and nutrition, but its prognostic role in critically ill ischemic stroke (IS) patients is unclear.Methods A cohort of 3,112 patients from MIMIC-IV was analyzed. WHR was calculated from first 24-hour ICU lab values. Its association with 28-day all-cause mortality(ACM) was assessed via survival analysis, Cox regression, restricted cubic splines (RCS), and subgroup analysis. Machine learning (ML) models were developed and evaluated using Receiver Operating Characteristic curve(AUC), calibration, and decision curve analysis.Results Mortality differed significantly across WHR quartiles (Log-rank p < 0.0001). A higher WHR was independently associated with increased 28-day ACM (adjusted HR per SD = 1.42; 95% CI: 1.22–1.65, p < 0.001), showing a near-linear dose-response. WHR (AUC = 0.644) outperformed its components (white blood cells, hemoglobin). The association was stronger in non-ventilated patients (interaction p = 0.003). Among ML models, the Gradient Boosting Machine (GBM) performed best (test-set AUC = 0.814) and WHR was a key predictive feature.Conclusion WHR is an independent predictor of short-term mortality in critically ill IS patients. Integrating WHR into ML models like GBM improves risk stratification, offering a valuable tool for clinical prognosis.