R2eD AVM Score A Novel Predictive Tool for Arteriovenous Malformation Presentation With Hemorrhage

Background and Purpose- The management of unruptured brain arteriovenous malformations remains unclear. Using a large cohort to determine risk factors predictive of hemorrhagic presentation of arteriovenous malformations, this study aims to develop a predictive tool that could guide hemorrhage risk stratification. Methods- A database of 789 arteriovenous malformation patients presenting to our institution between 1990 and 2017 was used. A hold-out method of model validation was used, whereby the data was randomly split in half into training and validation data sets. Factors significant at the univariable level in the training data set were used to construct a model based on multivariable logistic regression. Model performance was assessed using receiver operating curves on the training, validation, and complete data sets. The predictors and the complete data set were then used to derive a risk prediction formula and a practical scoring system, where every risk factor was worth 1 point except race, which was worth 2 points (total score varies from 0 to 6). The factors are summarized by R(2)eD arteriovenous malformation (acronym: R(2)eD AVM). Results- In 755 patients with complete data, 272 (36%) presented with hemorrhage. From the training data set, a model was derived containing the following risk factors: nonwhite race (odds ratio [OR]=1.8; P=0.02), small nidus size (OR=1.47; P=0.14), deep location (OR=2.3; P<0.01), single arterial feeder (OR=2.24; P<0.01), and exclusive deep venous drainage (OR=2.07; P=0.02). Area under the curve from receiver operating curve analysis was 0.702, 0.698, and 0.685 for the training, validation, and complete data sets, respectively. In the entire study population, the predicted probability of hemorrhagic presentation increased in a stepwise manner from 16% for patients with no risk factors (score of 0) to 78% for patients having all the risk factors (score of 6). Conclusions- The final model derived from this study can be used as a predictive tool that supplements clinical judgment and aids in patient counseling.