Predictive modeling and machine learning show poor performance of clinical, morphological, and hemodynamic parameters for small intracranial aneurysm rupture

Abstract

Abstract Small intracranial aneurysms (SIAs) (< 5 mm) are increasingly detected due to advanced imaging, but predicting rupture risk remains challenging. Rupture, though rare, can cause devastating subarachnoid hemorrhage. This study analyzed 141 SIAs (101 unruptured, 40 ruptured) using semi-automatic morphological analysis and high-resolution, image-based blood flow simulations from 3D rotational angiography. Advanced morphological and hemodynamic parameters were extracted, with clustering applied to address multicollinearity. Univariate logistic regression identified cluster representatives, and forward selection highlighted the maximum height, Neck inflow rate, and Non-sphericity index as rupture predictors, though only the latter two were significant. Clinical variables like age, sex, and comorbidities were also assessed but failed to predict rupture risk. The full model showed overfitting, with a pseudo-R2 of 0.142 on the training set but only 0.032 on the test set. A simplified model using just Neck inflow rate and Non-sphericity index performed similarly poorly (pseudo-R2 of 0.034). Multiple machine learning classifiers were evaluated, with similar performance across models, supporting the model-independence of the results. Overall, neither morphological, hemodynamic, nor clinical variables reliably predicted rupture risk, highlighting the limitations of current methods and underscoring the need for prospective studies and multimodal approaches that integrate imaging biomarkers and compare small and large aneurysms for better risk stratification.