Introduction
Intracranial aneurysms (IAs) are a significant public health concern, affecting approximately 3% of the general population. While many aneurysms remain stable throughout a patient's life, their rupture leads to subarachnoid hemorrhage (SAH), a devastating event characterized by high rates of morbidity and mortality. A particular challenge in clinical practice is the management of "small" aneurysms, typically defined as those 7 mm or less in diameter. Although these small IAs have a lower individual risk of rupture compared to larger ones, their high prevalence means they account for nearly 40% of all clinical rupture cases.
Clinicians currently face a difficult dilemma: whether to treat these small aneurysms through invasive surgical or endovascular procedures—which carry their own risks—or to opt for conservative monitoring. Traditional risk assessment tools, such as the PHASES score, often lack the precision required for small IAs because they rely on linear statistical models that may not capture the complex, non-linear interactions between various clinical and morphological risk factors.
Methodology
To address this gap, researchers conducted a retrospective study involving 1,004 patients with small IAs recruited from two different medical centers. The primary objective was to develop and validate a machine learning (ML) model capable of accurately predicting rupture risk using readily accessible clinical data.
The study compared five distinct ML algorithms: Support Vector Machine (SVM) with a linear kernel, Random Forest Classifier (RFC), Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). To ensure the model remained practical for clinical use, the researchers employed the Least Absolute Shrinkage and Selection Operator (LASSO) method to narrow down 21 initial variables to the 9 most predictive features.
The dataset was divided into a training and internal validation set (940 patients from one hospital) and an external validation set (64 patients from a second hospital) to test the model's generalizability. A critical component of the methodology was the use of Shapley Additive Explanations (SHAP) analysis, which provides a transparent look into the model's decision-making process by quantifying the contribution of each feature to the final risk score.
Results and Performance
The Support Vector Machine (SVM) emerged as the superior model, demonstrating both high performance and stability. In the internal validation cohort, the SVM achieved an Area Under the Curve (AUC) of 0.817 (95% CI, 0.769–0.866). More impressively, it maintained its predictive power in the external validation cohort with an AUC of 0.893 (95% CI, 0.808–0.979).
When compared to the traditional PHASES score, the ML models showed a significant improvement in predictive accuracy. The PHASES score only achieved AUCs of 0.616 and 0.667 in the internal and external cohorts, respectively. This disparity highlights the advantage of machine learning in identifying subtle patterns within clinical data that traditional scoring systems overlook.
Interpretability and Clinical Insights
The integration of SHAP analysis transformed the SVM from a "black box" into an interpretable clinical tool. The analysis identified the top three predictors of rupture as:
- Maximum Size: Even within the "small" category, incremental increases in size significantly impact risk.
- Location: Aneurysms located in the Anterior Communicating Artery (ACoA) and Internal Carotid Artery (ICA) were associated with higher rupture risk.
- Irregular Shape: Morphological irregularity served as a major "force multiplier" for risk.
An interesting finding was the "Hypertension Paradox," where patients with a documented history of hypertension appeared to have a lower risk of rupture in this cohort. The authors hypothesize that this may reflect a treatment effect, where patients aware of their condition are managed with medications that stabilize the aneurysm wall, whereas undiagnosed hypertension remains the true underlying threat.
Conclusion and Limitations
This study demonstrates that an interpretable SVM model can significantly improve the assessment of rupture risk for small intracranial aneurysms. By focusing on nine readily accessible features, the model provides a practical and transparent tool to assist clinicians in making informed treatment decisions.
However, the study is not without limitations. As a retrospective analysis, the morphological data was collected post-rupture for many cases, which might differ from the pre-rupture state. Additionally, the external validation cohort was relatively small. Future work should involve prospective, multicenter studies to further refine the model and incorporate more complex hemodynamic or radiomic data to potentially enhance predictive accuracy even further.
Summary updated