Background: Intracranial aneurysm (IA) is one of the most common causes of subarachnoid hemorrhage which is associated with high mortality and morbidity. Predicting aneurysms’ rupture risk has been an ongoing challenge clinically. The objective of this study is to utilize machine learning algorithms to predict IA’s rupture status based on simple parameters that can be measured in the daily clinical setting. Methods: A total of 448 IA cases with known rupture status (228 unruptured and 220 ruptured) were collected across three tertiary hospitals between November 2017 and November 2020. Twelve demographic and morphological parameters were obtained from all cases. Six machine learning models, namely Naive Bayes, Support Vector Machine, K-Nearest Neighbours, Gradient Boosting, Decision tree and Random Forest algorithms were applied to the dataset. Sensitivity, specificity and accuracy of each model was calculated along with Area Under the Curve (AUC) and F1 score. Furthermore, the relevant importance of all 12 variables was assessed. Results: Gradient Boosting and Random Forest algorithms were the ML algorithms with the highest AUC of 0.69, followed by Naive Bayes (AUC=0.66). Overall, Random Forest algorithm was the best-formed model with 67% sensitivity and 63% specificity. Support Vector Machine model gave the lowest favorable AUC of 0.45 with 45% sensitivity and 48% specificity. The most significant variable was the aneurysm neck width, followed by the parent artery diameter. Conclusions: Machine learning has the potential to assist with the IA’s rupture prediction by using simple variables that can be easily obtained in clinical practice. Further research, especially a larger sample size is required to improve diagnostic accuracy before its clinical translation.
Read full abstract