🤖 Generated with AI assistance · Does not replace individualized medical advice.
AI Act Transparency Notice
This content was generated with the assistance of artificial intelligence tools. While the information is derived from the provided scientific source, it has not been independently verified by a human medical professional. This post is for informational and educational purposes only and does not constitute medical advice, diagnosis, or treatment recommendations. Always consult with a qualified healthcare provider for personal medical concerns.
Predicting Incomplete Occlusion in Intracranial Aneurysms: A Multimodal Machine Learning Approach
Key Facts
| Metric | Value | Context |
|---|---|---|
| Primary Outcome | Incomplete occlusion post-EVT | Increases re-rupture and retreatment risks PMID: 42692877 |
| Model Performance (Internal Test) | ROC-AUC: 0.861 | Highest discrimination among tested algorithms PMID: 42692877 |
| External Validation | ROC-AUC: 0.835 | Demonstrates generalizability to external cohort PMID: 42692877 |
| Key Predictors | Ruptured status, Hypertension, CBF ratio | Identified via SHAP analysis PMID: 42692877 |
Why It Matters / Context
Incomplete occlusion following endovascular treatment (EVT) of intracranial aneurysms (IAs) is a significant clinical challenge, directly increasing the risks of re-rupture and the need for retreatment. Traditional assessment methods often rely on subjective angiographic grading, which can lack consistency and predictive power for long-term outcomes. This study addresses this gap by developing an interpretable, multimodal machine learning model that integrates quantitative digital subtraction angiography (QDSA) hemodynamic features with radiomics. By moving beyond simple morphological assessment to include hemodynamic and texture-based imaging features, the model aims to provide a more robust, individualized risk estimate. The development of a web-based decision-support calculator further enhances clinical applicability, allowing for risk-stratified surveillance planning without requiring complex local computational infrastructure. PMID: 42692877
Details
The study employed a rigorous dual-center design, comprising an internal retrospective cohort (n = 1212), a prospective cohort (n = 246), and an external validation cohort (n = 327). The internal cohort was randomly divided into training and internal test sets at an 8:2 ratio. The modeling process included feature selection, Synthetic Minority Over-sampling Technique (SMOTE), five-fold cross-validation, and hyperparameter optimization, all restricted to the training data to prevent data leakage. Eight different machine learning algorithms were compared, with the multimodal random forest model emerging as the top performer. The incremental value of multimodal feature integration was statistically 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) were utilized to ensure model interpretability.
| Cohort | ROC-AUC | PR-AUC | Sensitivity | Specificity | Brier Score |
|---|---|---|---|---|---|
| Internal Test | 0.861 PMID: 42692877 | 0.589 PMID: 42692877 | 0.550–0.683 PMID: 42692877 | 0.847–0.881 PMID: 42692877 | 0.117–0.148 PMID: 42692877 |
| Prospective | 0.853 PMID: 42692877 | 0.657 PMID: 42692877 | 0.550–0.683 PMID: 42692877 | 0.847–0.881 PMID: 42692877 | 0.117–0.148 PMID: 42692877 |
| External | 0.835 PMID: 42692877 | 0.439 PMID: 42692877 | 0.550–0.683 PMID: 42692877 | 0.847–0.881 PMID: 42692877 | 0.117–0.148 PMID: 42692877 |
Note: Sensitivity and Specificity ranges represent the overall performance across the three cohorts as reported in the abstract. Brier scores indicate low overall prediction error.
The multimodal model consistently outperformed both the clinical-only and clinical-morphological models across all cohorts. Decision curve analysis demonstrated a positive net benefit, suggesting clinical utility. SHAP analysis identified ruptured status, hypertension, cerebral blood flow ratio, Radscore, and treatment modality as the most important predictors. The final model was implemented as a web-based decision-support calculator, which accepts manually entered clinical data and precomputed imaging-derived variables to generate individualized risk estimates. PMID: 42692877
Limitations
Critical Considerations:
* Retrospective Internal Cohort: The primary training data was retrospective, which may introduce selection bias and confounding variables not fully accounted for by the model.
* External Cohort Size: The external validation cohort (n = 327) is relatively small compared to the internal cohort (n = 1212), potentially limiting the statistical power of the external validation.
* Manual Data Entry: The web-based calculator requires manual entry of clinical and precomputed imaging variables, which introduces a risk of human error and limits real-time integration with hospital information systems.
* Generalizability to Other Populations: The study was conducted in a dual-center setting; performance in other geographic or demographic populations may vary.
* Lack of Prospective Randomized Trial: The model has not been tested in a prospective randomized controlled trial to definitively prove that its use improves patient outcomes compared to standard care. PMID: 42692877
Clinical Takeaway
Clinicians can utilize the web-based calculator to integrate QDSA and radiomics features for individualized risk stratification of incomplete occlusion, thereby optimizing surveillance intervals and reducing unnecessary retreatment in high-risk patients.
Full Reference
Hu SQ, Yu J, Turhon M, Liu J, Liu T, Chen R. Predicting Incomplete Occlusion Following Endovascular Treatment of Intracranial Aneurysms: Development and Validation of a Multidimensional Machine Learning Model. Academic Radiology. 2026. PMID: 42692877 [DOI: 10.1016/j.acra.2026.08.050]