Prediction of rupture risk of intracranial aneurysms using machine learning models based on clinico-radiomic features

HU Xiao-long; DENG Peng; TANG Xiao-yu; MA Ming; QIAN Jin-hong; WU Gang; CHENG Zhi-qi; GONG Yu-hui; WU Jian-dong; DING Zhi-liang

doi:10.13798/j.issn.1009-153X.2023.09.002

Abstract：Objective To explore the effectiveness of machine learning models based on clinico-radiomics features in predicting the risk of intracranial aneurysm rupture.Methods The clinical data of 153 patients with intracranial aneurysms admitted to our hospital from January 2019 to December 2022 were retrospectively analyzed.Univariate analysis was used to screen the clinical features affecting aneurysm rupture.DICOM format image data were collected,and the 3D Slicer software was used to reconstruct and segment the parent artery.The radiomics plug-in was used to extract the morphological features of the aneurysm,and the LASSO algorithm was used to screen the important morphological features affecting aneurysm rupture.Machine learning models based on clinical and morphological features were constructed,and the area under the curve(AUC),accuracy,precision,sensitivity,and specificity of each model in predicting the risk of intracranial aneurysm rupture were calculated.Results Of these 153 patients,43 patients had ruptured intracranial aneurysms(ruptured group)and 110 had unruptured intracranial aneurysms(unruptured group).Compared with the unruptured group,the proportion of patients with hypertension was significantly higher(P<0.05),the elongation and sphericity of the aneurysm morphological features were significantly reduced(P<0.05),while the minimum axial diameter,maximum axial diameter,maximum coronal diameter,maximum three-dimensional diameter,grid volume,surface area,surface area volume ratio,and voxel volume of the aneurysm were significantly increased in the ruptured group(P<0.05).The AUC of SVM,KNN and LR models based on the six morphological features were 0.73(95%CI 0.47～0.98),0.80(95%CI 0.72～0.87),0.75(95%CI 0.61～0.89),respectively.The AUC of SVM1,KNN1 and LR1 models based on the clinical feature of hypertension and morphological features were 0.79(95%CI 0.67～0.93),0.85(95%CI 0.79～0.92),0.83(95%CI 0.72～0.95),respectively.Conclusions The morphological features of intracranial aneurysms were automatically extracted based on radiomics technology,and six machine learning models were constructed,which can accurately identify the status of aneurysms,so as to intervene as early as possible for intracranial aneurysms with high risk of rupture,which has important clinical significance.