| Objective: Intracranial aneurysm(IA)is one of the common cerebrovascular diseases.Once ruptured,the mortality and disability rate are high.Therefore,it is very important to predict the risk of rupture of IA accurately and early.The purpose of our study is to build a variety of machine learning models to predict the risk of early rupture of IA by comprehensively screening the relevant factors affecting the stability of IA,and provide guidance for the treatment optimization and diagnosis and treatment plan of clinical IA patients.Methods: The patients who were diagnosed as IA by CT angiography and clinicians from January 2010 to June 2022 in a tertiary hospital in Dalian were collected,and the case data were collected,including the clinical characteristics,blood biochemical and blood routine test indicators,and IA morphological parameters.The factors affecting the stability of IA were analyzed through inter-group comparison and single-factor logistic regression;Screening relevant factors affecting the stability of IA into the prediction model,including random forest(RF),support vector machine(SVM)and artificial neural network(ANN)combines three machine learning methods with three level models(model level 1: based on the clinical characteristics of patients;model level 2:based on the clinical characteristics of patients+blood biochemical and blood routine detection indicators;model level 3: based on the clinical characteristics of patients+blood biochemical and blood routine detection indicators+aneurysm-related parameters).Sensitivity,specificity,accuracy and receiver operating characteristic(ROC)curve were used to evaluate and compare the performance of the model.Results: A total of 989 patients with IA were included in this study,including 561 patients in the stable group and 428 patients with aneurysm in the unstable group.Univariate analysis showed age,sex,hyperlipidemia,diabetes,coronary heart disease,cerebral infarction,presence of clinical symptoms,diastolic blood pressure,systolic blood pressure,whether blood pressure was maintained well,high-density lipoprotein,triglyceride,glucose,apolipoprotein A1,apolipoprotein B,lipoprotein,uric acid,serum cystatin C,white blood cell count,hemoglobin,neutrophil count,maximum height,neck diameter,aspect ratio,undulation index,nonsphericity index,height-width ratio,irregular shape,acus and location of aneurysms were important factors affecting the stability of intracranial aneurysms(P<0.1).Among the three level models constructed based on RF method,the sensitivity,specificity,accuracy and area under ROC curve of the training set in model level 1 are 72.8%,76.9%,75.1% and 0.748(0.719-0.778)respectively,and the test set is 71.9%,75.0%,73.6% and 0.734(0.688-0.781)respectively;The sensitivity,specificity,accuracy and area under the ROC curve of the training set at model level 2 were 83.7%,84.2%,84.0% and 0.839(0.814-0.864)respectively,and the test set was 76.3%,84.0%,80.6% and 0.801(0.760-0.843)respectively;The sensitivity,specificity,accuracy and area under the ROC curve of the three training sets at the model level were 80.5%,86.9%,84.1% and 0.837(0.812-0.863)respectively,and the test set was 78.8%,83.0%,81.1% and 0.809(0.768-0.850)respectively.Among the three level models constructed based on SVM,the sensitivity,specificity,accuracy and area under the ROC curve of the model level 1 training set are66.0%,76.5%,71.7% and 0.712(0.682-0.743)respectively,and the test set is 44.2%,63.4%,57.9% and 0.699(0.651-0.747)respectively;The sensitivity,specificity,accuracy and area under the ROC curve of the training set at model level 2 were 80.2%,80.0%,80.0% and 0.913(0.884-0.924)respectively,and the test set was 75.7%,84.4%,80.3% and 0.804(0.763-0.845)respectively;The sensitivity,specificity,accuracy,and area under the ROC curve of the three training sets at the model level were 78.7%,85.5%,82.3%,and 0.824(0.798-0.849)respectively,and the test sets were 78.3%,83.0%,80.9%,and 0.806(0.765-0.848)respectively;Among the three level models constructed based on ANN method,the sensitivity,specificity,accuracy and area under the ROC curve of the model level 1 training set are 66.8%,75.3%,71.6% and 0.783(0.757-0.808)respectively,and the test set is 63.1%,65.5%,64.4% and 0.680(0.593-0.694)respectively;The sensitivity,specificity,accuracy and area under the ROC curve of the training set at model level 2 were 83.0%,79.6%,81.1% and 0.897(0.879-0.914)respectively,and the test set was 74.4%,79.0%,76.7% and 0.826(0.784-0.869)respectively;The sensitivity,specificity,accuracy,and area under the ROC curve of the three training sets at the model level were 81.9%,82.3%,82.1%,and0.897(0.879-914),respectively,and the test sets were 76.3%,84.0%,80.6%,and 0.860(0.821-0.899),respectively.With the increase of characteristic variables included,the prediction performance of RF,SVM and ANN models is gradually enhanced,and no significant difference is found among the three methods in the training set;The results of cross validation show that in model level 1,the prediction performance of RF is significantly better than that of SVM and ANN(P=0.032),and no significant difference is found at other model levels.The results of variable importance showed that age,white blood cell count and uric acid played an important role in predicting the stability of IA.Conclusion: The prediction model based on three machine learning methods combined the clinical characteristics of patients,blood biochemical and blood routine indicators,and aneurysm-related parameters has good performance in IA stability evaluation.The patient’s age,WBC and UA play an important role in predicting the stability of intracranial aneurysms as potential important predictors.In clinical treatment,we should pay attention to the changes of these three characteristic variables to prevent adverse outcomes. |
