P.126 Prediction of cerebral vasospasm using radiographical and clinical features: a machine learning model

Abstract

Background: Cerebral vasospasm is a feature of delayed cerebral ischemia that can occur after subarachnoid hemorrhage from cerebral aneurysm rupture. CTa is the universal first line diagnostic modality (sensitivity 79.6%, specificity 93.1%). We aim to determine whether specific vasospasm-associated radiographical and clinical features predicts cerebral vasospasm with comparable accuracy. Methods: Our cohort included 403 patients between 2006-2019. We used clinical predictive features including: day since rupture, transcranial doppler Lindegaard ratio, MCA velocity, ICA velocity, physical examination, and radiographical predictive features including: volume of hematoma, artifact, aneurysm, as our training dataset with true positives being digital subtraction angiography confirmed vasospasm. We used a decision-tree classifier from Scikit-learn library for training and testing of the model. Results: Our model trained on clinical and radiographical predictive features achieved sensitivity 93%, specificity 67%, F1 score 0.88. When using only radiographical features, we reached sensitivity 90%, specificity 55%, F1 score 0.83. When using only clinical features, we reach sensitivity 70%, specificity 93%, F1 score 0.87. Conclusions: We show that our vasospasm predictive model achieves adequate sensitivity, specificity, and F1 scores when compared to CTa. With further increase in dataset and fine-tuning of hyper-parameters, it is possible that our model may be used to optimize the vasospasm management pipeline.