Machine Learning on Routinely Monitored ICU Data Predicts Debilitating Cerebral Vasospasm Following Subarachnoid Hemorrhage in a Large Cohort

Introduction: Cerebral vasospasm is a feared complication occurring in 20-40% of patients following subarachnoid hemorrhage (SAH) and is known to contribute to delayed cerebral ischemia. It is standard practice to admit SAH patients to intensive care for several weeks of vigilant, resource-intensive, clinical monitoring. We used machine learning for early prediction of CV requiring verapamil in the largest known published cohort to date.

Methods: SAH patients admitted to UCLA from 2013-2021 were included. The UCLA perioperative data warehouse was reviewed for patient demographics and standardized clinical ICU data including vitals, routine labs, intracranial pressure, oxygenation/ventilatory information, fluid status, and feeding method until the primary endpoint, namely first verapamil administration or ICU downgrade. A light gradient boosting machine was trained to predict the primary endpoint using five-fold cross validation at various prediction timepoints. Receiver-operator curves (ROC) and precision-recall curves were generated at each prediction timepoint.

Results: A total of 1,750 patients were included, 125 of whom received verapamil. Average areas under the ROC curve were 0.93, 0.91, 0.90, and 0.88 at 4 hours, 1 day, 2 days, and 3 days, respectively, prior to the primary endpoint. Through 3 days prior, average precision was 0.09 at recall=1.00, and average recall was 0.27 at precision=1.00 for “verapamil given”. Age, minimum and average leukocyte count, and highest platelet count contributed most to predictive outcomes.

Conclusion : We present a highly accurate (average AUC>0.9) and early (up to 3 days prior) predictor of cerebral vasospasm warranting verapamil over the largest published cohort to date. This model allows for rule-out of 9% of patients not requiring verapamil (without any false negatives), and rule-in of 27% of patients requiring verapamil (without any false positives), averaged over each prediction timepoint. This work serves as a basis for optimized clinical management, improved resource allocation, and potential initiation of earlier ICU downgrade in SAH.