Machine Learning-Assisted Prediction of Persistent Incomplete Occlusion in Intracranial Aneurysms From Angiographic Parametric Imaging-Derived Features.

Rationale And Objectives: To develop machine-learning (ML) models incorporating angiographic parametric imaging (API)-derived parameters in predicting persistent incomplete occlusion of intracranial aneurysms (IAs) after flow diverter (FD) treatment.

Materials And Methods: We analyzed 763 patients with 834 FD-treated IAs, split into training (n=583) and test (n=251) sets. The training set was further subdivided into incomplete (n=96) and complete occlusion (n=487) groups according to O'Kelly-Marotta grading. We developed six ML models using clinical, morphological, and API-derived features identified via univariate analysis. We assessed model performance using the area under the curve (AUC), accuracy, and calibration curves. We used Shapley additive explanation (SHAP) analysis to highlight key predictive features and subgroup analysis with Benjamini-Hochberg correction to examine the impact of FD plus coiling.

Results: The Random Forest model outperformed all other models in predicting persistent incomplete occlusion after FD treatment (AUC=0.803 and accuracy=0.833 in the test set). SHAP analysis identified cerebral blood flow (CBF) ratio, radiomics scores (Radscore), incorporated branches, age, hypertension, and FD plus coiling as critical predictors. Multiple factors were significant in the FD-alone subgroup, whereas only recurrent aneurysms, higher CBF ratio, and elevated Radscore remained significant in the FD-plus-coiling subgroup. A CBF ratio cutoff of 0.49 was identified for IAs treated with FD plus coiling.

Conclusion: ML models that integrate clinical, morphological, and API-derived features effectively predict persistent incomplete occlusion after FD treatment. Key predictors included CBF ratio, Radscore, incorporated branches, age, hypertension, and FD plus coiling.