To develop and validate machine learning models to enable early prediction of sepsis in patients with aneurysmal subarachnoid hemorrhage (aSAH), with the goal of improving risk stratification and guiding timely interventions.

aSAH is a neurological emergency with well established high morbidity and mortality. Sepsis complicates 21–46% of aSAH cases and is independently associated with increased mortality, prolonged ICU stay, and worse functional outcomes. However, conventional tools such as SOFA and SIRS scores are limited in this population due to overlapping clinical features—such as neurogenic fever, sedation, and vasospasm—that mimic sepsis and reduce diagnostic specificity. Up to 85% of aSAH patients meet SIRS criteria irrespective of infection. Recent studies highlight the need for tailored prediction tools, with emerging evidence suggesting that individualized models outperform traditional criteria. The MIMIC-IV ICU database offers granular, longitudinal data on demographics, vital signs, laboratory results, and interventions, creating an opportunity to train and validate robust machine learning models. 

Machine learning models were trained on 803 adult patients with aSAH from the MIMIC-IV ICU database, including first ICU admissions >24 hours and excluding non-aneurysmal SAH and pregnancy/lactation. Thirty-four admission-time variables—including demographics, vital signs, laboratory values, ventilation status, patient history, and other key features—were used to predict sepsis (Sepsis-3) via 10-fold cross-validation in Python and WEKA. The models were subsequently externally validated on 763 adult patients using the eICU Collaborative Research Database cohort with identical features.

Across 10-fold cross-validation, the Random Forest model (ROC-AUC = 0.915) and a convolutional neural network (Multilayer Perceptron, ROC-AUC = 0.868) achieved the best discrimination for sepsis prediction. External validation using the eICU dataset confirmed generalizability (Random Forest = 0.833, MLP = 0.820).

These findings represent the first externally validated AI approach for sepsis prediction in aneurysmal subarachnoid hemorrhage. Future work will incorporate imaging features and prospective validation to enable bedside neuro-ICU deployment.