Machine Learning-based CT Imaging Radiomics Predicts Risk Of Perivalvular Leak After TAVR In Patients With Trilobular Aortic Valve

Background:Calcific Aortic Valve Disease(CAVD)is a group of degenerative diseases,mainly presenting as aortic stenosis,that occur in the elderly.Surgical Aortic Valve Replacement(SAVR)and Transcatheter Aortic Valve Replacement(TAVR)are the main treatment options for this group of diseases.The common postoperative complications of TAVR include stroke,perivalvular leak,vascular complications,conduction block,etc.Perivalvular leak is one of the most important,which directly affects the patient’s postoperative quality of life and can lead to heart failure or even death in severe cases.In the course of TAVR surgery,we found that the occurrence of postoperative perivalvular leakage correlates with the patient’s aortic root anatomy,and the observation of the patient’s preoperative aortic root structure by coronary CTA is one of the common preoperative evaluation methods for TAVR.Radiomics is an emerging research field in recent years,based on machine learning to analyze CT,MR,ultrasound and other imaging data,which can perform deeper information mining compared with conventional image analysis and has some advantages over traditional image analysis.Artificial intelligence has developed rapidly in recent years,and machine learning based on artificial intelligence can process complex data by simulating human learning in a computer,which is often used to solve radiomics problems.Objective:To investigate the relationship between CT imaging and postoperative perivalvular leak in trilobar aortic valve patients with TAVR by means of radiomics machine learning,to explore the CT imaging risk factors for postoperative perivalvular leak in this group of patients,to establish a CT imaging-based postoperative perivalvular leak prediction model and to test the accuracy of the model.Methods:A retrospective cohort study collected 180 patients with trilobular aortic valve(31 with postoperative perivalvular leak and 149 without perivalvular leak)who underwent transcatheter aortic valve placement at Qilu Hospital of Shandong University from March 2017 to December 2022.Baseline data and coronary CTA image data were collected,and the correlation between baseline data and postoperative perivalvular leak was statistically analyzed.Subsequently,the CTA images were segmented,and three regions of aortic valve calcification,valve noncalcification,and sinus were outlined as regions of interest and reconstructed as 3D models.The 3D image features of the region of interest were extracted,and some of the 2D image features in the CT plane commonly used for TAVR evaluation were mapped manually.Finally,the extracted features were analyzed by computer,and the most correlated image features were selected from the 3D image features of the calcified part of the valve,the 3D image features of the noncalcified part of the valve,the 3D image features of the sinus part of the valve,and the 2D image features using principal component analysis,variance selection,t-test selection,and the least absolute shrinkage selection algorithm LASSO,respectively.Finally,four models,including logistic regression,random forest,support vector machine,and plain Bayes,were used to predict postoperative perivalvular leakage and verify the efficiency,followed by plotting ROC curves to evaluate the efficacy of the models and select the prediction model with the best prediction.Results:1.p-values of all baseline data features in patients with perivalvular leak compared with non-perivalvular leak patients were greater than 0.05,with no statistical difference.2.1854 features were extracted from the patients’ CT images,and 16 image features with strong correlation with postoperative perivalvular leak were obtained by machine learning feature screening,such as the mean density of calcified valves and the degree of regularity of sinus shape,including 6 3D image features of calcified valves,6 3D image features of non-calcified valves,1 3D image feature of sinus,and 3 2D image features.3.Prediction models were trained using patient image features,and multiple perivalvular leak prediction models were developed and saved for valve calcification part 3D image features,valve noncalcification part 3D image features,sinus part 3D image features,and all image features,respectively.The model types with the best prediction results for the four groups of features were random forest model,plain Bayesian model,plain Bayesian model,and random forest model,and the validation accuracy and AUC were 87.8%,0.935;85.7%,0.914;98.0%,0.876;and 85.7%,0.85,in that order.the best prediction was the random forest model based on the 3D image features of the calcified part of the valve.Conclusion:1.The occurrence of perivalvular leakage after TAVR in patients with trilobar aortic stenosis is closely related to the anatomic characteristics of the patient’s aortic root,valve,and left ventricular outflow tract,and not to general clinical information.2.Machine learning CT radiomics is a valid method for preoperative evaluation of TAVR and can be used to assess the anatomic characteristics of patients with aortic stenosis and the risk of postoperative perivalvular leak.3.Factors that suggest an increased risk of postoperative perivalvular leak in patients with trilobular aortic valves based on machine-learning CT imaging include:(1)the more widespread the distribution of large calcifications in the valve,the greater the likelihood of perivalvular leak,whereas smaller punctate calcifications are less influential in perivalvular leak;(2)the higher the mean density of valve calcifications and the harder the calcifications,the greater the likelihood of perivalvular leak;(3)the thinner the thinnest part of the valve,the greater the likelihood of perivalvular leak(4)the more irregular and less spherical the sinus portion of the valve,the greater the likelihood of perivalvular leak;(5)the wider the left ventricular outflow tract,the greater the likelihood of perivalvular leak;(6)the greater the number of preoperative junctional fusion of the valve,the greater the likelihood of perivalvular leak.4.A predictive model of CT imaging based on machine learning can well predict perivalvular leakage after TAVR in patients with trilobular aortic valve.