Predicting Aneurysm Occlusion After Pipeline Embolization: an Ensemble Model Using Angiographic Parametric Imaging

BACKGROUND AND PURPOSE: The Pipeline Embolization Device (PED) revolutionized endovascular treatment of intracranial aneurysms (IAs), especially complex wide-neck lesions, but incomplete occlusion remains common. This study aimed to develop an ensemble model integrating clinical, morphological, and hemodynamic features for predicting post-PED occlusion outcomes. MATERIALS AND METHODS: This retrospective study included 306 patients with IAs treated with PED from January 2016 to June 2023. Clinical, morphological, and hemodynamic features were extracted from clinical records, three-dimensional rotational angiography, and digital subtraction angiography (DSA) to develop a weighted soft-voting ensemble model for predicting aneurysm occlusion at ≥6-month follow-up. The final prediction was generated by combining the weighted predicted probabilities from five individual algorithms rather than by a simple majority vote. Performance used area under the receiver operating characteristic curve (AUC) and confusion-matrix metrics; feature importance used SHapley Additive exPlanations (SHAP). RESULTS: Among 306 patients (mean age 52.9±10.9 years; 192 women), pre- and post-PED hemodynamic features were extracted; 31 features (13 clinical/morphological, 18 post-PED hemodynamic-9 each from aneurysm and parent-artery regions of interest) were modeled. Complete occlusion was observed in 202/243 aneurysms (83.1%) in the internal dataset and 49/63 (77.8%) in the external test set. In the external test set, the ensemble achieved an AUC of 0.902 (95% CI, 0.82-0.98), an accuracy of 88.9% (95% CI, 78.5-95.2%), a sensitivity of 93.9% (95% CI, 82.8-98.7%), and a specificity of 71.4% (95% CI, 45.0-88.3%). Independent predictors of incomplete occlusion were branch involvement, daughter sacs, and older age. Post-PED changes (time-to-peak, mean-transit-time, cerebral-blood-volume) correlated with outcomes; complete occlusion associated with higher aneurysmal maximum intensity projection and parent-artery cerebral blood flow. CONCLUSIONS: DSA-derived quantitative hemodynamic and morphological features are valuable biomarkers for PED efficacy. The ensemble showed robust performance (AUC 0.902; 95% CI 0.82-0.98) and may aid risk stratification and personalized IA treatment planning.