Clinical features are the primary measures used for risk assessment of cerebrovascular diseases. However, clinical features, especially angioarchitecture, in digital subtraction angiography require further interpretation by specialized radiologists. This approach for risk assessment requires multivariable analysis and is, therefore, challenging when completed manually. In this study, we employed three machine learning models, namely the random forest, naïve Bayes classifier, and support vector machine, for the detection of hemorrhagic brain arteriovenous malformations using digital subtraction angiography. Quantitative measurements from digital subtraction angiography were used as features, and the chi-squared test, minimum redundancy maximum relevance, ReliefF, and two-sample t tests were used for feature selection. Bayesian optimization was conducted to optimize the hyperparameters of the three models. The random forest model outperformed the other two models. As a human control, three radiologists diagnosed an independent testing data set. The random forest model had a computation time of less than a second for the whole data set for classification. Accuracy and the area under the receiver operating characteristic curve were 92.7% and 0.98 for the training data set and 85.7% and 0.97 for the independent testing data set, respectively. Compared with the mean diagnosis time of approximately half a minute per patient and the highest accuracy of 76.2% for the three radiologists, the random forest model was faster and more accurate for our data set. These results suggest that the machine learning model based on hemodynamic features from quantitative digital subtraction angiography is a promising tool for detecting hemorrhagic brain arteriovenous malformations.
- Brain arteriovenous malformation
- Digital subtraction angiography
- Machine learning
- Quantitative analysis
- Rupture risk
- Time-density curve