Admission laboratory parameters for early risk stratification after aneurysmal subarachnoid hemorrhage: a machine learning analysis of real-world data.

BACKGROUND: Aneurysmal subarachnoid hemorrhage (aSAH) is associated with substantial morbidity and mortality. Early risk stratification is essential for treatment planning and prognostication, yet neurological assessment may be limited in sedated or critically ill patients. Routine admission laboratory parameters may provide additional prognostic information and improve machine learning (ML)-based outcome prediction. RESEARCH QUESTION: To evaluate the predictive value of routinely obtained admission laboratory parameters for unfavorable neurological outcome after aSAH and to compare different ML algorithms. MATERIAL AND METHODS: This retrospective analysis included prospectively collected data from 389 patients with aSAH. Unfavorable outcome was defined as modified Rankin Scale (mRS) > 2 at 6 months. Four feature sets were analyzed: laboratory values, clinical variables, radiographic features and a combined feature set. Logistic regression, random forest, XGBoost and support vector machine models were evaluated using 5-fold-stratified-cross-validation. Feature importance was assessed using SHAP analysis. RESULTS: Laboratory-only models achieved moderate discrimination, with random forest demonstrating the highest performance (AUC = 0.724). Clinical models showed superior predictive accuracy overall (best AUC = 0.785), whereas radiographic-only models achieved lower performance. The combined feature set achieved the highest discrimination and outperformed the clinical-only model (p = 0.016), whereas laboratory-only and clinical-only discrimination did not differ significantly. The most important laboratory predictors were red blood cell distribution width, leukocyte count, glucose, and C-reactive protein. DISCUSSION AND CONCLUSION: Routine admission laboratory parameters carry modest but incremental prognostic information beyond established clinical grading when integrated through non-linear ML, but they do not replace clinical assessment. These single-center findings are hypothesis-generating and require external validation before clinical use.