Computational fluid dynamics and shape analysis enhance aneurysm rupture risk stratification

The management of patients with unruptured intracranial aneurysms (UIA) remains a clinical challenge. While 3% of adults have one, only a small percentage of these rupture [1]. When they do rupture, they cause subarachnoid haemorrhage, which results in significant morbidity and mortality [2].

Currently, the only options for prophylactic treatment are endovascular and surgical interventions [2]. Despite advances in these approaches, they remain associated with complication rates ranging from 3 to 10% [3, 4]. Subsequently, in clinical practice, the decision on whether to treat an aneurysm relies on comparing the perceived risk of its rupture against the risks associated with its treatment. In recent years, a number of natural history studies have investigated factors associated with aneurysm rupture. The standard risk stratification model for estimating the risk of rupture is the PHASES score [5], which uses six clinical variables (ethnicity, hypertension, age, aneurysm size, aneurysm location, history of previous SAH) to estimate the five-year risk. While this scoring system achieved an AUC (Area Under the Curve, where the curve is the Receiver Operating Characteristic, ROC) of 0.82 on its original dataset, its performance deteriorates when applied to external data [6‐8].

Although the best available model, the PHASES score remains a crude assessment of the risk of rupture and there remains a need for more detailed and personalised risk estimates. Angiographic imaging and derived haemodynamic and morphological data hold vast potential. In recent years, the development of computational fluid dynamics (CFD) has provided insight into aneurysm rupture, with factors like wall shear stress [9], oscillatory shear index [10] and residence time [11] already explored. This has led to the understanding that aneurysms with irregular shapes (e.g. high aspect ratio [12] or presence of bulges [13]) or that exhibit complex flow patterns [14] are more prone to rupture, and identification of novel risk factors derived from shape and haemodynamic analyses. These factors have been integrated into several rupture classification models [14‐17], although none has been implemented in clinical practice. One reason for this is that CFD and morphology models have not been tested to see if they add any predictive value over and above routinely available clinical variables. To improve their clinical uptake, it is also necessary to reduce their lengthy processing, feature-extraction, and evaluation times, to show that the feature extraction operation can be performed repeatably and reliably, and to show that the adoption of computer-derived data is beneficial in a clinical context.

This study’s first aim is to evaluate whether haemodynamic and morphological data generated within the pipeline developed as part of the European project @neurIST [18], can improve the distinction between ruptured and unruptured aneurysms compared to the PHASES score. Its second aim is to identify which image-derived features of aneurysm shape and flow most strongly correlate with rupture status in order to optimise the feature set and facilitating integration of computational risk models with existing clinical workflows.

This study assessed five logistic regression models constructed from different combinations of clinical, morphological (shape), and haemodynamic data for classifying aneurysms’ rupture status. Previous studies have identified several morphological and haemodynamic factors which show statistical association with aneurysm rupture, and have shown that they can be used to classify ruptured and unruptured cases [7, 14, and 26]. However, they have not assessed whether the addition of these can improve established risk prediction models with respect to clinically established protocols. Our results show that, when compared to the PHASES score alone, the addition of either morphological or haemodynamic data individually improved classification accuracy. The addition of both morphological and haemodynamic data did not cause any further improvement.

Throughout our study, we also showed how through statistical approaches on a relatively large dataset we can reduce the number of image-derived features to those that demonstrate strongest association. This has been extensively debated through the scientific community [27‐29], which recognises the importance of using robust statistical approaches and considerations when evaluating new potential image-derived features. This helps to avoid the proliferation of potentially confounding and misleading features, often arising from small, single-centre studies with limited populations. Such features may lack statistical significance and hinder generalisability of results.

Our results show that haemodynamic analysis can effectively improve the separation between ruptured and unruptured cases. We have found that high OSI and low WSS are indicative of rupture, which is in accordance with previous observations [9, 26, and 30]. In the final set of retained variables, WSS does not directly appear through its values, but rather in terms of the area of the region where its values are high or low. This derives from the feature-reduction stage of the analysis, where we observed that MaxWSSPeak showed a high correlation (r = 0.88, r < 0.001) with MaxVPeak and was thus removed from the pool. The extension of aneurysm surface area exposed to low WSS (AbsLowWSSAreaAvg) is positively associated with rupture. This implies that, in case of rupture, large parts of the aneurysm are exposed to low WSS, a condition that has been identified as responsible for initiation, growth and rupture because of their role in endothelial remodelling [10, 31]. The association of WSS with rupture, however, is controversial and there exist studies that have linked rupture with high WSS [32]. This inconsistency can be partially explained by the observations in [33], where the authors hypothesised that aneurysms of different size fail because of various combinations of high/low WSS and oscillatory shear index (OSI). The OSI is a quantity derived from WSS that quantifies the oscillatory behaviour of the WSS vector along the cardiac cycle [34]. High OSI values are indicative of complex and dynamic flow patterns, vortices and recirculation regions within the aneurysm which ultimately induce low WSS and lower residence time [35]. Our models identify these conditions as associated with rupture status, coherently with results reported in [9, 36].

Complex haemodynamics is often induced by irregular shapes and geometries. NSI and AR are generally recognised as aggregated metrics for describing aneurysms’ shape, with higher values describing shapes that differ from an ideal sphere [12, 37]. Deviations from spherical shape yield uneven distributions of wall stress which enhance aneurysm instability and favour rupture [12]. NSI higher than 0.2 have shown significant discriminatory power in multiple studies [12, 37]. Our cohort of ruptured patients presented NSI = 0.198 ± 0.07, and that feature ranked as highly influential in our analysis. AR was part of the initial set of variables but, despite being widely acknowledged as a risk factor for aneurysm rupture [12, 37], it was excluded from the final analysis. This was because of its correlation with the PHASES variable AneuDepth (r = 0.79, p < 0.001) and, by design, we decided to retain all the PHASES variables. Studies [5, 38] observed that patients with ruptured aneurysms showed higher values of AneuDepth. This association is confirmed in our analysis when only the PHASES variables are used as predictors but show an opposite direction when derived data are included. Previous studies [39, 40] have shown that narrow necks and larger volumes can induce blood stagnation within the aneurysm, thus creating haemodynamic conditions favourable for rupture. We did not observe this phenomenon in our results, and we ascribe it to the features of our cohort, where patients with large volumes and narrow necks tend to belong to the unruptured group. Other researchers have reported conclusions similar to ours [15, 41].

Despite this study showing the improved classification of aneurysm rupture with the addition of CFD and morphological derived factors, there remain several limitations. In terms of the CFD, the boundary conditions were obtained from generic 1D models of the brain circulations, while personalised models might offer a more precise representation of the boundary conditions. In terms of modelling, we did not include in our analysis variables derived from patients’ lifestyle such as smoking and diet, or related to familial history of ruptured aneurysms and other cardiovascular pathologies. These are very well-established risk factors and regularly taken into consideration by doctors when evaluating possible treatment strategies [42]. However, they are not included among the PHASES variables and, since our study aims at evaluating the potential of CFD and shape analysis alongside the use of PHASES variables, we ultimately decided to exclude them from our model.

In terms of the population, our study is similar to previous studies [10, 26], and suffers from limited sample size and the cross-sectional nature of the dataset. The cross-sectional design of this study can explain a number of findings. It is not clear what is the effect of rupture on aneurysm shape and volume. Some authors report increases or while others report no modification [43, 44] in small-cohort studies. Our models identify smaller aneurysms (both in terms of Vol and AneuDepth) as predictive of unruptured status: however, without longitudinal observations to track their evolution, it is not possible to reach definitive conclusions on their role. Additionally, age showed an ambiguous effect in our models with elderly patients being more likely to rupture when exclusively considering the PHASES features, and less prone to rupture when shape and haemodynamics are taken into account. While there are conflicting views regarding the role of age [45], elderly patients are more likely to have a diagnosis of an UIA and this will influence the structure of a dataset. Regarding hypertension, it was not associated with rupture. Our dataset did not include any actual blood pressure values, however. It is therefore possible that more patients with ruptured aneurysms had undiagnosed hypertension, whereas more patients with unruptured aneurysms had their blood pressure controlled on antihypertensive agents. Furthermore, only 27% of the patients in the ruptured group suffered from hypertension, which likely introduced a bias. Our dataset also included a similar number of ruptured and unruptured cases and did not represent the distribution of aneurysms in the general populations. This is an issue common to most currently available studies [10, 13, 15]. The dataset was also cross sectional and was used to develop a classifier capable of distinguishing someone with a ruptured aneurysm from someone with an unruptured aneurysm. The dataset did not allow us to assess if it could predict which unruptured aneurysms would go on to rupture in the future, which is the clinical question that needs to be addressed and what PHASES was developed for. Practically compiling the necessary longitudinal datasets to answer this question is challenging, however, due to the relatively low short rates of rupture, which mean very large datasets with very long periods of follow up are necessary. This is now being addressed in the Risk of Aneurysm Rupture (ROAR) study which is following up the largest cohort of patients with unruptured aneurysms to date with more than double the cases of the whole PHASES metanalysis combined and far longer periods of follow up available [46]. Finally, our model was only validated internally, and we did not perform external validation. This is common to most computational studies [14‐16, 21, 24], with the notable exception of [47]. The dataset used in this study, however, were collected during the @neurIST project which, despite being a large multicentric effort involving twenty-nine partners from twelve countries, did not have a uniformed data collection protocol. This implicitly guarantees that our model is robust to various images modalities. The size of the dataset we used, 170 patients, did not allow for further subdivisions based on the hospitals, while at the same time maintaining a meaningful sample size. Further validation of this model can be performed by resorting, for example, to larger datasets such as AneuX [48].

In conclusion, we showed that using additional data derived from CFD and morphological analysis increases the ability of logistic regression models in separating ruptured aneurysms from unruptured ones using clinical variables alone. The resulting logistic regression models achieved AUC = 0.71 and used a reduced number of features which were obtained through semi-automated processing of medical images and CFD results. This approach has the potential to be included within current clinical protocols, once being extended and validated using longitudinal data and larger patient cohorts.