Shared and Distinct Rupture Discriminants of Small and Large Intracranial Aneurysms
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.
The most widely used surrogate for assessing intracranial aneurysm (IA) rupture risk is aneurysm size.1–3 This metric was adopted from longitudinal prospective studies, which found that larger aneurysms are more likely to rupture.1,3,4 However, studies have shown that 13% to 75% of ruptured aneurysms are small.5–8 This poses a dilemma for treatment decisions of small IAs. Furthermore, there is no consensus on the definition of small and large IAs7: 5,9,10 7,1,3,4 and 10 mm2 have all been used as cutoffs.
Clinical evidence suggests that small and large IAs tend to have different pathophysiological presentations.11–14 Intraoperative observations show that smaller IAs (<4 mm) typically have thin, translucent walls, whereas larger IAs (>10 mm) tend to have thick, irregular, and whitish walls.11 Histological studies of ruptured and unruptured IAs have found that small aneurysms (<10 mm) are characterized by thin, hypocellular walls, whereas large IAs (>10 mm) have thick atherosclerotic walls with the presence of inflammatory cells.11,12 These differences in the IA wall anatomy suggest that small and large IAs may need to be analyzed separately.
On the other hand, because size may not adequately reflect the risk of IA rupture,15–17 alternative rupture characteristics have been investigated for rupture discrimination models.17,18 However, these models were built without establishing a cutoff measure to distinguish small and large aneurysms. We hypothesized that small and large IAs could have different morphological, hemodynamic, and clinical characteristics associated with rupture, and such differences may affect image-based rupture discrimination models.
In this study, we investigated whether building rupture classification models separately for small and large IAs could better discriminate IA rupture status. We performed a cross-sectional analysis of morphological, hemodynamic, and clinical rupture characteristics of both small and large IAs in an existing data set. To define small and large aneurysms objectively, we used unsupervised hierarchal cluster analysis to determine an inherent size cutoff that would dichotomize the IA population. Multivariate logistic regression was used to build rupture discrimination models based on small and large IAs separately and in combination (aggregate model). Last, we validated the 3 new models and determined the performance of the model predictions in an independent testing cohort.
Materials and Methods
The data that support the findings of this study are available from the corresponding author on reasonable request. The University at Buffalo Institutional Review Board approved the protocol for this study, which waived the need for patient consent. All patients were from Gates Vascular Institute. The inclusion criteria included: patients who had 3-dimensional digital subtraction or computed tomographic angiography, a confirmed ruptured or unruptured IA, and sufficient image quality for segmentation with no noticeable artifacts to accurately represent the IA and surrounding vasculature. The accurate geometry is necessary for computational fluid dynamics. Two cohorts of patients with IAs were retrospectively collected: (1) the training cohort, consisting of patients evaluated between May 2006 and May 2013 (n=413), and (2) the testing cohort, consisting of patients evaluated between June 2015 and May 2016 (n=129). The IAs in the training cohort were comprehensively analyzed for morphological, hemodynamic, and clinical features and were used in building rupture discrimination models. The testing cohort was used for model validation.
Analyses of Morphological, Hemodynamic, and Clinical Features
We computed 3-dimensional morphological and hemodynamic features from segmented images (Table 1; analyses and features defined in the online-only Data Supplement and Table I in the online-only Data Supplement). Image segmentation and computational fluid dynamics methods were described previously.15,17,19–21 To gain further insight into the flow dynamics of IAs, the inflow was classified as either Jet Breakdown Mode or Continuous Jet Mode.22 We also collected and analyzed clinical features, including previous subarachnoid hemorrhage (SAH), hypertension, and IA location (Table 2).
Size Cutoff Defining Small and Large Aneurysms
To determine if small and large IAs intrinsically separate, hierarchal cluster analysis was performed on the training cohort. Morphological and hemodynamic features, excluding IA size, determined clustering. All features were standardized on a scale from 0 to 1. To assess the ability of an IA size cutoff to determine cluster assignment, we built a receiver operating characteristic (ROC) curve. If the ROC analysis showed good performance (area under the curve [AUC]>0.7), it would be determined that separating the population into 2 groups based on IA size is reasonable. The size cutoff would be determined by the Youden Index (J),23 which gives equal weight to sensitivity and specificity.
Construction of Small, Large, and Aggregate Rupture Discrimination Models
To select candidate morphological, hemodynamic, and clinical features for rupture discrimination models, we first compared ruptured and unruptured IAs by univariate analyses in the small IA, large IA, and aggregate groups. Binary features were compared by χ2 or Fisher exact tests; continuous features were first tested for normality, followed by a Student t test (for normally distributed data) or a Mann–Whitney U test (for non-normally distributed data). Statistically significant features (P<0.05) were considered for logistic regression analysis. Features were also tested for collinearity and were considered for logistic regression if they were independent (P>0.05; correlation coefficient <0.8).
To build rupture discrimination models for small IAs, large IAs, and the aggregate group, we performed multivariate logistic regression. Using a backwards conditional stepwise method, a single rupture discrimination model was built for each group that included morphological, hemodynamic, and clinical features of IAs. Continuous morphological and hemodynamic characteristics were scaled to a range of 0 to 10 to ease the comparison of odds ratios (ORs). Binary factors were included in the equation as absent (0) or present (1). Statistical analyses were performed using IBM SPSS Statistics Version 22.0 (IBM Corporation, Armonk, NY).
Assessment of Model Performances
We evaluated the performance of the size-dichotomized and aggregate models in the training cohort by comparing the AUCs that were the result of the logistic regression analysis. We predicted that an IA was ruptured when the probability was greater than the cost-effective cutoff (sensitivity weighted twice more than specificity).24
To test the ability of the models to predict rupture status in an independent cohort of IAs, we applied them to the testing cohort. We treated the small and large IA models as a unified, size-dichotomized model and compared it to the aggregate model. The AUC, sensitivity, specificity, and accuracy (percentage of IAs with correctly predicted rupture status) of each model were determined.
Figure 1 shows the distributions of IAs by size and location in the training and testing cohorts. The training cohort consisted of 413 IAs (102 ruptured) from 345 patients with an average IA size of 4.84±3.26 mm. The testing cohort consisted of 129 IAs (14 ruptured) from 107 patients with an average size of 4.92±2.95 mm.
Size Cutoff Defining Small and Large Aneurysms
Hierarchal cluster analysis shows that the 413 IAs separated into 2 primary clusters (cluster A and cluster B; Figure 2A). Ellipticity index, or the deviation of the IA convex hull from a perfect hemisphere, was the most important feature for determining cluster assignment (Figure 2B). Cluster A tended to have IAs that were more hemispherical and smaller (Figure 2C). Further, ROC analysis revealed that size was a good determinant of cluster assignment (AUC=0.81) and the best size cutoff to separate small and large IAs was 5.04 mm (Figure 2D). Similar size cutoffs were found when ruptured and unruptured IAs were analyzed independently (5.24 and 4.55 mm; Figure in the online-only Data Supplement). Therefore, 5 mm was adopted as a cutoff to dichotomize the training cohort into 269 small (63 ruptured) and 144 large (39 ruptured) IAs. When the morphological and hemodynamic features of the 2 groups were compared, 13 of 16 features were statistically different (Table II in the online-only Data Supplement).
Rupture Characteristics of Small and Large Aneurysms
Tables 1 and 2 show descriptive statistics from the univariate comparison of ruptured and unruptured IAs for small (<5 mm) and large (≥5 mm) IAs in the training cohort. Small and large IAs had several similar rupture discriminants. In both groups, ruptured IAs had higher size ratio, undulation index (UI), Ellipticity index, nonsphericity index, oscillatory shear index (OSI), relative residence time, low wall shear stress area, and lower normalized wall shear stress (WSS), compared with unruptured IAs. In addition, ruptured IAs more often had Jet Breakdown Mode, occurred more frequently in patients with previous SAH and at the anterior cerebral artery location, and occurred less frequently at the internal carotid artery location.
Small and large IAs have some different rupture discriminants. Small ruptured IAs had higher aneurysm number and were less frequent in the presence of multiple IAs when compared with small unruptured IAs. Large ruptured IAs had a higher maximum WSS, occurred in younger patients, and a higher frequency at the posterior communicating artery location than large unruptured IAs. The results of univariate analysis for the aggregate group are provided in Tables III through V in the online-only Data Supplement.
Rupture Discrimination Models
Multivariate logistic regression analysis yielded 3 rupture discrimination models for small IAs, large IAs, and the aggregate group:(1)(2)(3)
where odds is the ratio of the probability of ruptured status (P) to the probability of unruptured status (1−P) of a given IA.
Independently significant discriminants of small IA rupture included UI, OSI, previous SAH, and the absence of multiple IAs. Discriminants of large IAs included UI, low WSS, previous SAH, and anterior cerebral artery or posterior communicating artery location. The aggregate model retained all previously mentioned small and large IA rupture discriminants.
Figure 3 shows examples of ruptured and unruptured small and large IAs with statistics for the 269 small and 144 large IAs in the training cohort. We found that both small and large IA rupture discrimination models contain UI, or the degree of IA surface irregularity. We also found that high OSI predicted rupture in small IAs, whereas low WSS predicted rupture in large IAs.
For all 3 models, previous SAH had a higher OR than any other model parameter (small IAs OR, 3.86; 95% confidence interval [CI], 1.29–11.51; large IAs OR, 7.65; 95% CI, 1.28–45.69; aggregate OR, 9.55; 95% CI, 3.61–25.23). The ORs for model parameters are given in Table V in the online-only Data Supplement.
Assessment of Rupture Discrimination Models
The ROC curves for the size-dichotomized and the aggregate models for small and large IAs in the training cohort were compared (Figure 4A and 4B). For small IAs, the AUC of the size-dichotomized model was marginally higher compared with the aggregate model (0.84; 95% CI, 0.78–0.88 versus 0.82; 95% CI, 0.77–0.88). For large IAs, the AUC of the size-dichotomized model was also marginally higher when compared with the aggregate model (0.87; 95% CI, 0.82–0.93 versus 0.83; 95% CI, 0.75–0.90). From the ROC curves, we predicted that IAs would be ruptured at a probability >0.23, 0.34, and 0.21 for the small, large, and aggregate models, respectively.
Last, the performance of the size-dichotomized and aggregate models in the independently analyzed testing cohort was assessed (Figure 4C and 4D). The size-dichotomized model has higher sensitivity (64% versus 29%) and accuracy (77% versus 74%), marginally higher AUC (0.75; 95% CI, 0.61–0.88 versus 0.67; 95% CI, 0.52–0.82), and similar specificity (78% versus 80%) compared with the aggregate-based model.
Identifying the risk of IA rupture is challenging. Many longitudinal and cross-sectional studies have attempted to identify rupture risk predictors, with size being the most widely accepted metric. In the PHASES study (Population, Hypertension, Age, Size, Earlier subarachnoid hemorrhage, and IA Site),1 pooled analysis of 6 longitudinal studies including ISUIA (International Study of Unruptured Intracranial Aneurysms)3 and UCAS (Unruptured Cerebral Aneurysm Study)4 found that the estimated risk of aneurysm rupture increases with aneurysm size. For IAs between 7 and 10 mm, the risk is 2.7× that of small IAs (<5 mm); and for those >20 mm, the risk increases to 14.3×.1 This suggests that small IAs are least likely to rupture. However, clinical reports show that a substantial percentage of ruptured aneurysms are small. In some reports, 35%8 or 47%5 of ruptured IAs are <5 mm; in another report, 66% of ruptured IAs are <7 mm.6 The disparity between the reported low rupture risk and high rupture presentation of small IAs illustrates a gap in our understanding of small IA rupture.
The high percentage of ruptured small IAs among ruptured IAs clearly indicates that small IAs can rupture without becoming large IAs, but we do not know how long it has taken for them to reach the point of rupture. Given that longitudinal studies, such as those surveyed by the PHASES study,1 tend to follow low-risk IAs that are slated for observation and any sign of increased risk would usually result in treatment, findings from these studies may not represent the true rupture risk of small IAs. In this light, there are some merits in cross-sectional studies that include both small and large ruptured IAs.15–17,19,21,25
Most previous studies, including both longitudinal and cross-sectional ones, have examined IAs without dichotomizing them by size, but small and large IAs could be different phenotypes or even different pathologies.11–14 In our study, hierarchal cluster analysis resulted in 2 primary IA groups separated at 5 mm in size. This suggests that there are some intrinsic differences between small and large IAs in morphological and hemodynamic features. We found that large IAs (≥5 mm) were more irregular in shape with lower WSS, and small IAs (<5 mm) were more spherical with higher WSS.
We inquired whether small and large IAs might have different rupture risk profiles. We found that they have some shared and some distinct rupture discriminants. The shared discriminants include (1) a history of SAH, consistent with previous studies,1,3,9 and (2) higher UI (surface irregularity), and this is also in line with previous studies.9,19,25,26 Irregular shapes may reflect the presence of daughter aneurysms. These geometries tend to harbor secondary vortices which may promote inflammatory cell infiltration.17 This type of local disturbed flow environment may also be conducive to thrombus formation, which can exacerbate inflammatory cell infiltration and increase the proteolytic degradation of the IA wall.12,14,27
The distinct discriminants of small and large IAs include their hemodynamic environment. Although low WSS predicted large IA rupture, high OSI predicted small IA rupture. Meng et al14 proposed 2 distinct hemodynamic-biological pathways that are responsible for IA growth and rupture: low WSS and high OSI in the aneurysm sac could trigger inflammatory cell–mediated destructive remodeling, whereas a high-flow condition could trigger mural cell–mediated destructive remodeling. Furthermore, it was proposed that the dominance of the inflammatory cell–mediated pathway could give rise to large, irregular, and atherosclerotic IAs and the dominance of a mural cell–mediated pathway could be responsible for the genesis of small, smooth-walled IAs.14
In our study, we found that low WSS predicted rupture status in large IAs. This could indicate the dominance of the inflammatory cell–mediated pathway in these IAs. On the other hand, high OSI predicted rupture status in small IAs in our study. Meng et al14 suggested that jet impingement on the wall could lead to growth and rupture in small IAs. In our study, we found that Jet Breakdown Mode, which is a result of an impingement jet, was associated with rupture. To determine if Jet Breakdown Mode was associated with high OSI, we performed a post hoc analysis on the training cohort. Our analysis revealed that OSI was significantly higher in IAs with Jet Breakdown Mode than in IAs with Continuous Jet Mode (Table VI in the online-only Data Supplement). This may be because the breakdown of an inflow jet creates a more unsteady flow at the aneurysm wall, thus resulting in high OSI. We also suspect that high OSI plays a bigger role in small IAs compared with large IAs because of a relatively larger surface area being affected by jet impingement.
In addition to hemodynamics, small and large IAs also had distinct rupture discriminants in IA locations and multiplicity of IAs. Large IAs had a higher frequency of rupture at anterior cerebral artery and posterior communicating artery locations, as in previous studies.1,4,26 In small IAs, IA multiplicity was negatively associated with rupture status. This is consistent with previous findings that in the presence of multiple IAs, the ruptured one is most likely the largest aneurysm.28
This study has several limitations. First, our patient population was from a single center, and we do not know how applicable the results are to different populations. Second, we adopted several commonly used assumptions to make computational fluid dynamics tractable. Because of a lack of patient-specific information, we assumed a generic inlet waveform and a constant, location-based inlet flow rate. The inlet velocities were scaled by the inlet diameter. We also assumed blood as a Newtonian fluid and that IAs have rigid walls. Last, our rupture discrimination models were based on cross-sectional data, and we cannot address whether they can predict impending rupture in unruptured IAs. Specifically, evidence suggests that the postrupture IA geometry may be different from the prerupture geometry.29,30 Therefore, the observed phenomenon may represent the change in an IA geometry secondary to rupture.
In this study of 542 IAs, hierarchal cluster analysis showed that small and large IAs fell into 2 clusters based on morphological and hemodynamic features, with 5 mm being the best cutoff. Small (<5 mm) and large (≥5 mm) IAs have different hemodynamic and clinical, but not morphological, rupture discriminants. Furthermore, size-dichotomized rupture discrimination models based on morphological, hemodynamic, and clinical parameters had a higher sensitivity and accuracy than such a model built from the aggregate cohort with no size distinction. These results may indicate that small and large IAs have different paths to rupture.
N. Varble and Dr Meng conceived and designed the article; N. Varble, Dr Sonig, Dr Siddiqui, and Dr Davies acquired data; N. Varble, V.M. Tutino, Dr Yu, and Dr Davies analyzed and interpreted data; all authors drafted, critically revised, and approved the article.
We thank Dr Jianping Xiang for assistance with computational analysis, Paul H. Dressel BFA for preparation of the illustrations, and Elaine C. Mosher MLS and Debra J Zimmer for editorial assistance.
Sources of Funding
This work was supported by National Institutes of Health grants R01NS091075 and R03NS090193, Toshiba Medical Systems Corp, and resources from the Center for Computational Research at the University at Buffalo.
V.M. Tutino is a cofounder of Neurovascular Diagnostics, Inc. Dr Yu is a coinvestigator of National Institutes of Health (NIH) grant R01NS091075. Dr Siddiqui received financial support from Apama Medical, Buffalo Technology Partners, Inc, Cardinal, Endostream Medical, Ltd, International Medical Distribution Partners, Medina Medical Systems, Neuro Technology Investors, StimMed, and Valor Medical; is a consultant of Amnis Therapeutics, Ltd, Cerebrotech Medical Systems, Inc, Cerenovus (Formerly Codman Neurovascular, Neuravi, and Pulsar Vascular), CereVasc, LLC, Claret Medical, Inc, Corindus, Inc, GuidePoint Global Consulting, Integra (Formerly Codman Neurosurgery), Medtronic (Formerly Covidien), MicroVention, Neuravi (Now Cerenovus), Penumbra, Pulsar Vascular (Now Cerenovus), Rapid Medical, Rebound Therapeutics Corporation, Silk Road Medical, Stryker, The Stroke Project, Inc, Three Rivers Medical, Inc, Toshiba America Medical Systems, Inc, and W.L. Gore & Associates; is a member of advisory board of Intersocietal Accreditation Commission; and is a member of National Steering Committees of Codman & Shurtleff LARGE Trial, Covidien (Now Medtronic) SWIFT PRIME (Solitaire With the Intention for Thrombectomy as Primary Endovascular Treatment) and SWIFT DIRECT (Bridging Thrombolysis Versus Direct Mechanical Thrombectomy in Acute Ischemic Stroke) Trials, MicroVention FRED Trial (Flow-Redirection Endoluminal Device) & CONFIDENCE Study (Carotid Artery Stenosis in Patients at Elevated Risk for Adverse Events From Carotid Endarterectomy), MUSC POSITIVE Trial (Medical University of South Carolina Perfusion Imaging Selection of Ischemic Stroke Patients for Endovascular Therapy), Penumbra 3-Dimensional Separator Trial, COMPASS Trial (Cardiovascular Outcomes for People Using Anticoagulation Strategies), INVEST Trial (International Verapamil SR-Trandolapril), Neuravi ARISE II Trial (Analysis of Revascularization in Ischemic Stroke With EmboTrap) Steering Committee. Dr Davies receives research grant from National Center for Advancing Translational Sciences KL2 grant; and is a member of speakers’ bureau of Penumbra; Honoraria: Neurotrauma Science, LLC. Dr Meng is a coinvestigator of NIH grant 1R01NS091075, R03NS090193. The other authors report no conflicts.
The online-only Data Supplement is available with this article at http://stroke.ahajournals.org/lookup/suppl/doi:10.1161/STROKEAHA.117.019929/-/DC1.
- Received October 31, 2017.
- Revision received January 29, 2018.
- Accepted February 1, 2018.
- © 2018 American Heart Association, Inc.
- Greving JP,
- Wermer MJ,
- Brown RD Jr,
- Morita A,
- Juvela S,
- Yonekura M,
- et al
- Rinkel GJ,
- Djibuti M,
- Algra A,
- van Gijn J
- Wiebers DO,
- Whisnant JP,
- Huston J III,
- Meissner I,
- Brown RD Jr,
- Piepgras DG,
- et al
- Lee GJ,
- Eom KS,
- Lee C,
- Kim DW,
- Kang SD
- Murayama Y,
- Takao H,
- Ishibashi T,
- Saguchi T,
- Ebara M,
- Yuki I,
- et al
- Wermer MJ,
- van der Schaaf IC,
- Algra A,
- Rinkel GJ
- Kadasi LM,
- Dent WC,
- Malek AM
- Kataoka K,
- Taneda M,
- Asai T,
- Kinoshita A,
- Ito M,
- Kuroda R
- Meng H,
- Tutino VM,
- Xiang J,
- Siddiqui A
- Xiang J,
- Natarajan SK,
- Tremmel M,
- Ma D,
- Mocco J,
- Hopkins LN,
- et al
- Cebral JR,
- Mut F,
- Weir J,
- Putman C
- Takao H,
- Murayama Y,
- Otsuka S,
- Qian Y,
- Mohamed A,
- Masuda S,
- et al
- Varble N,
- Trylesinski G,
- Xiang J,
- Snyder K,
- Meng H
- Lindgren AE,
- Koivisto T,
- Björkman J,
- von Und Zu Fraunberg M,
- Helin K,
- Jääskeläinen JE,
- et al
- Schneiders JJ,
- Marquering HA,
- van Ooij P,
- van den Berg R,
- Nederveen AJ,
- Verbaan D,
- et al
- Crompton MR
- Schneiders JJ,
- Marquering HA,
- van den Berg R,
- VanBavel E,
- Velthuis B,
- Rinkel GJ,
- et al