Shared and Distinct Rupture Discriminants of Small and Large Intracranial Aneurysms

Abstract

Background and purpose: Many ruptured intracranial aneurysms (IAs) are small. Clinical presentations suggest that small and large IAs could have different phenotypes. It is unknown if small and large IAs have different characteristics that discriminate rupture.

Methods: We analyzed morphological, hemodynamic, and clinical parameters of 413 retrospectively collected IAs (training cohort; 102 ruptured IAs). Hierarchal cluster analysis was performed to determine a size cutoff to dichotomize the IA population into small and large IAs. We applied multivariate logistic regression to build rupture discrimination models for small IAs, large IAs, and an aggregation of all IAs. We validated the ability of these 3 models to predict rupture status in a second, independently collected cohort of 129 IAs (testing cohort; 14 ruptured IAs).

Results: Hierarchal cluster analysis in the training cohort confirmed that small and large IAs are best separated at 5 mm based on morphological and hemodynamic features (area under the curve=0.81). For small IAs (<5 mm), the resulting rupture discrimination model included undulation index, oscillatory shear index, previous subarachnoid hemorrhage, and absence of multiple IAs (area under the curve=0.84; 95% confidence interval, 0.78-0.88), whereas for large IAs (≥5 mm), the model included undulation index, low wall shear stress, previous subarachnoid hemorrhage, and IA location (area under the curve=0.87; 95% confidence interval, 0.82-0.93). The model for the aggregated training cohort retained all the parameters in the size-dichotomized models. Results in the testing cohort showed that the size-dichotomized rupture discrimination model had higher sensitivity (64% versus 29%) and accuracy (77% versus 74%), marginally higher area under the curve (0.75; 95% confidence interval, 0.61-0.88 versus 0.67; 95% confidence interval, 0.52-0.82), and similar specificity (78% versus 80%) compared with the aggregate-based model.

Conclusions: Small (<5 mm) and large (≥5 mm) IAs have different hemodynamic and clinical, but not morphological, rupture discriminants. Size-dichotomized rupture discrimination models performed better than the aggregate model.

Keywords: hemodynamics; intracranial aneurysm; machine learning; rupture.

Figure 1

Distributions of all IAs in the training and testing cohorts based on (A) size and (B) location. Distribution of ruptured IAs based on (C) size and (D) location.

Figure 2

Results of hierarchal cluster analysis. (A) The IAs, labeled by size, and separated into two clusters. (B) EI was the most important feature in separating the two clusters. (C) A scatter plot of EI versus IA size indicated that more hemispherical and smaller IAs tended to be assigned to Cluster A. (D) ROC analysis showed that IA size was a good indicator of cluster assignment (AUC=0.81) and the best size cutoff to separate small and large IAs was 5.04mm. AR=aspect ratio; EI=ellipticity index; LSA=low wall shear stress area; NSI=nonsphericity index; RRT=relative residence time; SR=size ratio; UI=undulation index; WSS=wall shear stress

Figure 3

Representative unruptured and ruptured IAs with statistics for all 269 small and 144 large IAs in the training cohort. The aneurysm sac was highlighted to illustrate UI (significant for both small and large IAs) in (A) small IAs and (B) large IAs. (C) OSI distributions in small IAs. (D) Normalized WSS distributions in large IAs.

Figure 4

ROC analysis of size-dichotomized and aggregate models for (A) small IAs and (B) large IAs in the training cohort. (C) Model performance in the testing cohort showing the higher AUC of the size-dichotomized model compared to the aggregate model. (D) The size-dichotomized model also had higher sensitivity, and accuracy compared to the aggregate model.