Machine Learning Models for Brain Arteriovenous Malformations Presenting with Hemorrhage Based on Clinical and Angioarchitectural Characteristics.

RATIONALE AND OBJECTIVES: This study aims to develop the best diagnostic model for brain arteriovenous malformations (bAVMs) rupture by using machine learning (ML) algorithms.

MATERIALS AND METHODS: We retrospectively included 353 adult patients with ruptured and unruptured bAVMs. The clinical and radiological data on patients were collected. The significant variables were selected using univariable logistic regression. We constructed and compared the predictive models using five supervised ML algorithms, multivariable logistic regression, and R2eDAVM scoring system. A complementary systematic review and meta-analysis of studies was aggregated to explore the potential predictors for bAVMs rupture.

RESULTS: We found that a small nidus size of <3 cm, deep and infratentorial location, longer filling time, and deep and single venous drainage were associated with a higher risk of bAVMs rupture. The multilayer perceptron model showed the best performance with an area under the curve value of 0.736 (95% CI 0.67-0.801) and 0.713 (95% CI 0.647-0.779) in the training and test dataset, respectively. In our pooled analysis, we also found that the male sex, a single feeding artery, hypertension, non-White race, low Spetzler-Martin grade, and coexisting aneurysms are risk factors for bAVMs rupture.

CONCLUSION: This study further demonstrated the clinical and angioarchitectural characteristics in predicting bAVMs hemorrhage.