Predicting Incomplete Occlusion Following Endovascular Treatment of Intracranial Aneurysms: Development and
Sheng-Qi Hu1, Jiasheng Yu2, Mirzat Turhon3
1Tongji Junshan Neuroscience Center, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan City, Hubei Province, People's Republic of China (S.Q.H., J.Y., R.C.); Department of Interventional Neuroradiology, Beijing Neurosurgical Institute, Capital Medical University, Beijing, People's Republic of China (S.Q.H., M.T., J.L., T.L., W.L., X.Y.); Department of Neurosurgery, Beijing TianTan Hospital, Capital Medical University, Beijing, People's Republic of China (S.Q.H., M.T., J.L., T.L., W.L., X.Y.).
Rationale And Objectives:
Incomplete occlusion after endovascular treatment (EVT) of intracranial aneurysms (IAs) increases the risks of re-rupture and retreatment. We aimed to develop and validate an interpretable multimodal machine learning model integrating quantitative digital subtraction angiography (QDSA) and radiomics features.
Materials And Methods:
This dual-center study included an internal retrospective cohort (n = 1212), a prospective cohort (n = 246), and an external cohort (n = 327). The internal cohort was randomly divided into training and internal test sets at an 8:2 ratio. Clinical, morphological, QDSA-derived hemodynamic, and radiomics features were analyzed. Feature selection, SMOTE (Synthetic Minority Over-sampling Technique), five-fold cross-validation, and hyperparameter optimization were restricted to the training data. Eight algorithms were compared, and the incremental value of multimodal feature integration was assessed using DeLong tests. The primary performance metric was the area under the receiver operating characteristic curve (ROC-AUC), supplemented by the area under the precision-recall curve (PR-AUC), calibration measures, and decision curve analysis. Shapley Additive Explanations (SHAP) analysis was used for model interpretation, and the final model was implemented as a web-based decision-support calculator.
Results:
The multimodal random forest model achieved ROC-AUCs of 0.861, 0.853, and 0.835 and PR-AUCs of 0.589, 0.657, and 0.439 in the internal test, prospective, and external cohorts, respectively. Sensitivity ranged from 0.550 to 0.683 and specificity from 0.847 to 0.881. The model outperformed the clinical and clinical-morphological models across all cohorts and showed generally preserved discrimination across clinical subgroups. Overall prediction error remained low, with Brier scores ranging from 0.117 to 0.148, while decision curve analysis demonstrated positive net benefit. SHAP analysis identified ruptured status, hypertension, cerebral blood flow ratio, Radscore, and treatment modality as important predictors.
Conclusion:
The multimodal model demonstrated stable discrimination across three cohorts. The web-based calculator provides individualized risk estimates from manually entered clinical and precomputed imaging-derived variables and may support risk-stratified surveillance planning.