Machine learning prediction of imaging detected growth in unruptured intracranial aneurysms using clinical and morphological features

Purpose To develop and internally validate a machine learning model using clinical and morphological features to predict imaging detected growth of unruptured intracranial aneurysms (UIAs). Methods This single center retrospective study included 784 UIAs detected by non-invasive angiography between September 2008 and March 2021. A patient level 7:3 split created a training set and a patient independent test set. Seven algorithms were compared using leakage-free stratified 5×5 nested cross validation, with standardization, L1 penalized feature selection, and hyperparameter tuning repeated within the training folds. The selected model was refitted on the complete training set and evaluated once in the test set. Inverse probability of censoring weighted time dependent receiver operating characteristic curve analyses at 12 and 24 months compared the model with ELAPSS and PHASES. SHapley Additive exPlanations (SHAP) were used for interpretation, and a research only Streamlit prototype was developed. Results Among 784 UIAs from 613 patients, 257 (32.8%) had imaging detected growth, with a median follow-up of 375 days. The training set comprised 551 UIAs and the test set 233 UIAs. The support vector machine (SVM) had the highest mean outer-fold area under the curve (AUC) (0.875±0.013, 95% CI 0.858 to 0.891) and achieved a test set AUC of 0.881 (95% CI 0.835 to 0.923) and Brier score of 0.130 (95% CI 0.106 to 0.156). The time dependent AUC values of SVM were 0.772 (95% CI 0.686 to 0.844) at 12 months and 0.840 (95% CI 0.763 to 0.906) at 24 months. The non-sphericity index had the greatest mean absolute SHAP value. Conclusion Our SVM showed promising discrimination for imaging detected UIA growth. These exploratory findings require independent multicenter and prospective validation before clinical use.