Early Prediction of Cerebral Vasospasm After Aneurysmal Subarachnoid Hemorrhage Using a Machine Learning Model and Interactive Web Application

Abstract

Background: Cerebral vasospasm is a frequent and severe complication after aneurysmal subarachnoid hemorrhage (aSAH), often causing delayed cerebral ischemia (DCI) and poor outcomes. Despite progress in neurocritical care, early vasospasm prediction after aSAH remains challenging due to its multifactorial nature but is essential for timely intervention. Methods: We retrospectively analyzed 503 consecutive patients with spontaneous subarachnoid hemorrhage (SAH) treated between 2013 and 2018. Of these, 345 with angiographically confirmed aSAH were included in the primary analysis, and 158 SAH cases in a sensitivity analysis. We extracted demographic, clinical, and imaging parameters including age, sex, Hunt and Hess grade, Fisher scale, aneurysm and treatment features, external ventricular drainage (EVD), and central nervous system (CNS) infection. Seven supervised machine learning (ML) models, including logistic regression and gradient-boosted trees, were trained using nested cross-validation and evaluated by AUC-ROC, AUC-PR, accuracy, precision, sensitivity, specificity, and F1 score. Results: Over half of aSAH patients developed moderate to severe vasospasm. Independent predictors included younger age, higher Hunt and Hess and Fisher grades, and EVD placement (all p < 0.001). Logistic regression achieved the best discrimination (AUC-ROC 0.723), while tree-based models reached higher sensitivity (0.867) at the expense of specificity. Aneurysmal etiology further increased vasospasm risk (OR 4.72). Conclusions: Routinely available clinical and imaging parameters enable reliable ML-based vasospasm prediction after aSAH. Logistic regression provided the best balance between accuracy and interpretability, while tree-based models optimized sensitivity. This web-based, interpretable ML tool-one of the first using routine clinical data-may support the bedside prediction of vasospasm and requires prospective validation.

Keywords: aneurysmal subarachnoid hemorrhage; cerebral vasospasm; clinical decision support; machine learning; predictive modeling; risk stratification.

Figure 1

Bar charts illustrating the predictive performance of machine learning models in the aSAH cohort (N = 345) across nine evaluation metrics: (A) AUC-ROC, (B) AUC-PR, (C) AP, (D) PPV, (E) NPV, (F) bACC, (G) TPR, (H) TNR, and (I) F1-score. Each chart shows the mean metric values for all models, enabling visual comparison of their relative performance.

Figure 2

Precision–recall curves for all machine learning models predicting moderate or severe cerebral vasospasm: (a) in patients with aSAH; (b) in the full cohort. Curves were mean-aggregated across all test folds of repeated cross-validation, with the horizontal line indicating baseline precision. For each model, the point corresponding to the default decision threshold is indicated.

Figure 3

Receiver operating characteristic (ROC) curves for all machine learning models: (a) in patients with aSAH; (b) in the full cohort, predicting moderate or severe cerebral vasospasm. Curves were mean-aggregated across all test folds of repeated cross-validation, with the diagonal line representing baseline performance. For each model, the point corresponding to the default decision threshold is indicated.