Abstract Number ‐ 59: Study on Rupture Prediction Model for Small Unruptured Cerebral Aneurysms Using Random Forest

Abstract

Introduction Hemodynamic, morphological, and clinical factors have been reported to be involved in aneurysm rupture. Hemodynamic factors could be calculated by computational fluid dynamics (CFD) analysis. In addition, machine learning has been used in various medical fields in recent years, and it is expected to be applicable to the prediction of aneurysm rupture. In this study, we built a rupture prediction model for the small aneurysms (3∼10 mm) using Random Forest, which is one of the machine learning algorithms. This model included hemodynamic parameters from CFD simulations, morphological parameters of each aneurysm, and clinical information. The model was applied to the other patient dataset, and we verified the model’s prediction accuracy. Methods In this study, the ruptured case was defined as an aneurysm that ruptured during the follow‐up term. On the other hand, the unruptured case was defined as an aneurysm that remained stable during the follow‐up for more than two years (average follow‐up term is about nine years), and registered in the database of Jikei Hospital between January 1, 2003 and April 30, 2020. Only small aneurysms (3∼10 mm) were considered in the present study. Based on the criteria, we identified 507 aneurysms (ruptured: 41, unruptured: 466), in which CFD analysis has been completed. Among these aneurysms, 405 aneurysms (ruptured: 32, unruptured: 373) were used to build the rupture prediction model as Training data, and 102 aneurysms (ruptured: 9, unruptured: 93) were used to verify the prediction accuracy as Test data. We used 3‐dimensional arterial geometries that were reconstructed from computed tomography angiography images for conducting CFD simulations and morphological measurements. For the ruptured aneurysms, the image acquired before the rupture was used. We obtained 138 hemodynamic parameters, 7 morphological parameters, and 6 clinical information in each case. From the Training data, we proposed the rupture prediction model with these parameters using Random Forest. Then, we introduced the test data into the model, and its sensitivity and specificity were estimated. Results The sensitivity and specificity of this model for the Test data were 88.9% and 83.9%, respectively. In addition, aspect ratio (AR), the maximum height of the cerebral aneurysm (Hmax), and the spatially maximum values of the oscillatory shear index on the aneurysm wall (OSImax) were obtained as the top three of the important features to predict aneurysm rupture. Furthermore, the values of these parameters of ruptured aneurysms are higher than those of unruptured aneurysms. For unruptured and ruptured cases, average AR were 0.770±0.260 [‐] and 1.02±0.382 [‐], average Hmax were 3.39±1.22 [mm] and 4.96±1.96 [mm], and average OSImax were 0.399±0.105 [‐] and 0.460±0.0328 [‐], respectively. These results imply that cerebral aneurysms with high Hmax, high AR, and high OSImax are more likely to rupture. Conclusions Ruptured aneurysms tend to have high AR, high Hmax, and high OSImax. The sensitivity and specificity of our prediction model were 88.9% and 83.9%, respectively. The rupture prediction model obtained from this study may predict the aneurysm rupture in advance.