To utilize a machine learning and deep learning approach to derive a clinically applicable tool to predict the development of vasospasm in aneurysmal subarachnoid hemorrhage (aSAH) patients.
Cerebral vasospasm, is a common undesired and feared consequence of aSAH occurring in up to 70% of patients. Cerebral vasospasm may present anywhere from 24 hours to 7-10 days post SAH and is difficult to predict. Several radiological and clinical features have been proposed as potential predictors of cerebral vasospasm in previous models including, however, these predictors have not successfully been standardized and weighted through an artificial intelligence algorithm. While risk factors for vasospasm is known, there is also no validated single tool in practice that accounts for the broad diversity in indicators that may be predictive of cerebral vasospasm.
We retrospectively analyzed aneurysmal SAH patients who were admitted to the Carle Foundation Hospital in the Critical Care Unit from January 2016 to December 2021. Total of 52 patients met study criteria, out of which 23 (44%) patients were identified to have vasospasm through diagnostic imaging modalities. Clinical, radiological, and laboratory data were collected and analyzed through various feature selection and artificial intelligence models. Models were validated using 10-fold Cross-Validation.
We included Transcranial Doppler data from 14 days post aSAH, as well as Fisher Score, Hunt Hess Score, diameter of aneurysm, lymphocyte/monocyte/WBC counts, motor component from Glasgow Coma Scale, and home medication for anticoagulants in our model. Models including Logistic Model Trees, Logistic Regression, AdaBoost, LogitBoost, Hoeffding Tree, and Multilayer Perceptron (MP) were trained, with MP performing with the highest ROC AUC of 0.829, with a TP Rate of 73%, FP Rate of 28.5%, Precision of 73%, and F-Measure of 73%.
A framework for predicting cerebral vasospasm after aSAH through deep learning is created with ROC AUC of 0.829.