Using machine learning to predict the rupture risk of multiple intracranial aneurysms
Junqiang Feng1, Chunyi Wang1, Yu Wang1
1Department of Neurosurgery, Beijing Chaoyang Hospital, Capital Medical University, Beijing, China.
Objective:
This study aims to develop a machine learning-based risk prediction model (RPM) for the rupture of multiple intracranial aneurysms (MIAs), addressing a critical gap in current clinical tools such as the PHASES score, which are not specifically designed for MIAs. By analyzing detailed morphological and anatomical parameters, our model provides a tailored approach to rupture risk assessment in MIAs, offering potential improvements over existing methods.
Methods:
To address dataset imbalance, we conducted five-fold cross-validation. External validation was not feasible due to data limitations, but we rigorously evaluated model performance using metrics such as accuracy (ACC), true positive rate (TPR), true negative rate (TNR), F1 score, and area under the receiver operating characteristic curve (AUC).
Results:
Ninety-one patients with 222 aneurysms were recruited, with a rupture rate of 20.3%. The model demonstrated preferable predication performance in unruptured aneurysms (TNR: 0.837) but showed limitations in predicting ruptured aneurysms (TPR: 0.644). Error analysis revealed that the model's lower TPR may be attributed to the small sample size and dataset imbalance. Overall, the model achieved an accuracy of 0.797 and an AUC of 0.843.
Conclusion:
Our model provides a novel approach to predicting rupture risk in MIAs, complementing existing tools like the PHASES score. However, its clinical applicability is currently limited by suboptimal performance for ruptured aneurysms, which is more suited for identifying MIAs after rupture rather than predicting future rupture risk, and the lack of external validation. Future studies with larger, prospective cohorts are needed to validate and refine the model. This work highlights the potential of machine learning to enhance rupture risk assessment in MIAs, offering a foundation for more personalized treatment strategies.
Significance:
Multiple intracranial aneurysms have distinct mechanisms of formation, progression, and rupture. The widely used PHASES score does not incorporate morphological parameters of aneurysms and is not specifically designed for patients with multiple aneurysms. Therefore, we constructed a risk prediction model for the rupture of MIAs by machine learning algorithms.
Insights
This study developed a machine learning model to predict the rupture of multiple intracranial aneurysms (MIAs), outperforming existing scores for unruptured cases but needing refinement for ruptured aneurysms.
Area of Science:
- Neurosurgery
- Medical Imaging
- Machine Learning
Background:
- Multiple intracranial aneurysms (MIAs) present unique rupture risks not addressed by current tools like the PHASES score.
- Existing risk prediction models lack specific consideration for the morphological and anatomical parameters crucial in MIAs.
Purpose of the Study:
- To develop and evaluate a machine learning-based risk prediction model (RPM) for the rupture of MIAs.
- To provide a tailored risk assessment for MIAs, improving upon general clinical tools.
Main Methods:
- A five-fold cross-validation approach was used to manage dataset imbalance.
- Model performance was assessed using accuracy, true positive rate, true negative rate, F1 score, and AUC.
Main Results:
- The model achieved an overall accuracy of 0.797 and an AUC of 0.843.
- It demonstrated higher predictive performance for unruptured aneurysms (TNR: 0.837) compared to ruptured ones (TPR: 0.644).
- Lower TPR for ruptured aneurysms was linked to small sample size and data imbalance.
Conclusions:
- The developed ML model offers a novel approach for MIA rupture risk prediction, complementing the PHASES score.
- Clinical applicability is limited by suboptimal performance in predicting ruptured aneurysms and the absence of external validation.
- Future research with larger cohorts is essential to validate and enhance the model for personalized treatment strategies.
