Neural Network Helps Determine the Hemorrhagic Risk of Cerebral Arteriovenous Malformation

Abstract

We aimed to determine whether the hemorrhage risks of cerebral arteriovenous malformation (AVM), evaluated through digital subtraction angiography (DSA) using a neural network, were superior to those assessed through angioarchitecture. We conducted a retrospective review of patients with cerebral AVM who underwent DSA from 2011 to 2017. Angioarchitecture parameters, age, and sex were analyzed using univariate and multivariate logistic regression. Additionally, a neural network was trained using a combination of convolutional neural network (CNN) and recurrent neural network (RNN) architectures. The training dataset consisted of 118 samples, while 29 samples were reserved for testing. After adjusting for age at diagnosis and sex, single venous drainage (odds ratio [OR] = 2.48, p = 0.017), exclusive deep venous drainage (OR = 3.19, p = 0.005), and venous sac (OR = 0.43, p = 0.044) were identified as independent risk factors for hemorrhage. The angioarchitecture-based hemorrhagic prediction model achieved 69% accuracy with an AUC (area under the ROC curve) of 0.757, while the CNN–RNN-based model achieved 76% accuracy with an AUC of 0.748. We present a diagnostic performance for hemorrhagic risk assessment of AVMs that is comparable to the angioarchitectural analysis. By leveraging larger datasets, there is significant potential to enhance prediction accuracy further. The CNN–RNN algorithm not only can potentially streamline workflow within the angio-suite but also serves as a complementary approach to optimize diagnostic accuracy and treatment strategies.