A pilot study using a machine-learning approach of morphological and hemodynamic parameters for predicting aneurysms enhancement.

PURPOSE The development of straightforward classification methods is needed to identify unstable aneurysms and rupture risk for clinical use. In this study, we aim to investigate the relative importance of geometrical, hemodynamic and clinical risk factors represented by the PHASES score for predicting aneurysm wall enhancement using several machine-learning (ML) models. METHODS Nine different ML models were applied to 65 aneurysm cases with 24 predictor variables. ML models were optimized with the training set using tenfold cross-validation with five repeats with the area under the curve (AUC) as cost parameter. Models were validated using the test set. Accuracy being significantly higher (p < 0.05) than the non-information rate (NIR) was used as measure of performance. The relative importance of the predictor variables was determined from a subset of five ML models in which this information was available. RESULTS Best-performing ML model was based on gradient boosting (AUC = 0.98). Second best-performing model was based on generalized linear modeling (AUC = 0.80). The size ratio was determined as the dominant predictor for wall enhancement followed by the PHASES score and mean wall shear stress value at the aneurysm wall. Four ML models exhibited a statistically significant higher accuracy (0.79) than the NIR (0.58): random forests, generalized linear modeling, gradient boosting and linear discriminant analysis. CONCLUSIONS ML models are capable of predicting the relative importance of geometrical, hemodynamic and clinical parameters for aneurysm wall enhancement. Size ratio, PHASES score and mean wall shear stress value at the aneurysm wall are of highest importance when predicting wall enhancement in cerebral aneurysms.