| PartⅠ VALUE OF ECHOCARDIOGRAPHIC EPICARDIAL ADIPOSE AND MYOCARDIAL STRATIFICATION STRAIN PARAMETERS IN DIAGNOSING THE EXTENT OF CORONARY ARTERY LESIONS IN PATIENTS WITH CORONARY ARTERY DISEASEObjective: To analyze the diagnostic value of epicardial adipose tissue(EAT)thickness,stratification strain parameters including global longitudinal strain(GLS)and territorial longitudinal strain(TLS)measured by echocardiography on the extent of coronary artery lesions in patients with coronary artery disease(CAD).Methods:187 patients diagnosed with CAD by coronary angiography at the First Hospital of Lanzhou University Heart Center from September 2020 to November 2022 were included,and the degree of stenosis of the diseased coronary arteries was assessed with the help of Gensini score,and the participants were divided into 47 cases with mild(Gensini score≤24),53 cases with moderate(25≤Gensini score≤53)and severe group(Gensini score≥54)in 87 cases.All patients were enrolled for echocardiography followed by coronary angiography.The conventional echocardiographic parametersand end-systolic EAT thickness were measured;acquisition of GLS of the left ventricular endocardial,myocardial and epicardial layers(GLSendo,GLSmyo,GLSepi)and TLS of the left ventricular endocardial,myocardial and epicardial layers(TLSendo,TLSmyo,TLSepi)using the myocardial stratification strain technique.Results:1.The EAT of CAD patients in the mild,moderate and severe lesion groups showed a gradient thickening,and GLSendo,GLSmyo,GLSepi and TLSendo,TLSmyo,TLSepi showed a gradient decrease in that order(P<0.05).2.Multifactorial logistic regression analysis showed that EAT,GLSendo,and TLSendo were independent influencing factors for severe coronary artery lesions in patients with CAD(P<0.05).3.The area under curve(AUC)of subjects with EAT,GLSendo,and TLSendo predicting severe coronary artery lesions in patients with CAD were 0.714(0.641~0.787),0.819(0.757~0.881),and 0.788(0.724~0.852),respectively;cut-off values of 7.175 mm,-16.635,and-18.115,respectively;the sensitivity was 74.7%,69.0% and 85.1%,respectively;and the specificity was 63.0%,85.0% and 58.0%,respectively.4.Pearson correlation analysis showed a positive correlation between EAT thickness and Gensini score(r=0.340,P<0.001),GLSendo(r=-0.383,P<0.001),GLSmyo(r=-0.483,P<0.001),GLSepi(r=-0.436,P<0.001),TLSendo(r=-0.414,P<0.001),TLSmyo(r=-0.367,P<0.001),and TLSepi(r=-0.302,P<0.001)were negatively correlated with Gensini score.Conclusions:1.As the degree of coronary artery disease gradually increases in patients with CAD,the EAT gradually thickens and the overall and local systolic function of all layers of the myocardium gradually decreases.2.Myocardial stratified strain technique can rapidly quantify the overall and local function of the left ventricle in CAD patients,and EAT,GLS and TLS can be used as important references for predicting the severity of coronary artery lesions in CAD patients and identifying complex CAD non-invasively.PartⅡ CONSTRUCTION OF A PREDICTIVE MODEL FOR SEVERITY OF CORONARY ARTERY LESIONS INPATIENTS WITH CORONARY ARTERY DISEASE BASED ON MACHINE LEARNING ALGORITHM OF ULTRASOUND PARAMETERSObjective: To investigate the feasibility of constructing a machine learning model for the severity of coronary artery lesions in CAD patients based on the characteristic parameters of echocardiography and to test the efficacy of the prediction model.Methods:One hundred and eighty-seven patients diagnosed with CAD by coronary angiography at the First Hospital of Lanzhou University Heart Center from September 2020 to November 2022 were included,and the degree of stenosis of the diseased coronary arteries was assessed with the help of Gensini score.The participants were divided into 100 cases of mild to moderate(Gensini score≤53)and 87 cases of severe group(Gensini score≥54),and clinical and ultrasound parameters were collected.Stratified sampling was randomly performed by computer according to the severity of coronary lesions,with 75% of patients as the training set(n=140)and 25% as the test set(n=47).The best features of the prediction models were screened in the training set using the random forest(RF)algorithm,which were incorporated into six machine learning algorithms: decision tree(DT),K-nearest neighbor(KNN),logistic regression(LR),naive bayes(NB),RF,and extreme gradient boosting algorithm(XGBoost)for prediction model building,while the models were trained and internally validated using 3-fold cross-validation.The best parameters of each model are finally obtained,and the prediction performance of each machine learning model is verified in an independent test set.The effectiveness of each model is evaluated by AUC;the accuracy,sensitivity or recall,specificity,positive prediction value or precision,negative prediction value,and F1 score are used to evaluate the prediction classification effect of each model;after plotting the calibration curve of each model,the difference between the predicted and actual values of the model is analyzed,and the Brier score is further calculated to compare the overall performance of the predicted models,and the best model is selected as the final model.Results:1.The body mass index of patients in the training set was higher than that of patients in the test set(P<0.05);the differences in diastolic blood pressure,left ventricular ejection fraction,EAT,GLSendo,GLSmyo,GLSepi and TLSendo,TLSmyo,TLSepi were statistically significant(P<0.05)in patients in the mild to moderate andsevere lesion groups in the training set.2.After screening by RF algorithm,7 features of EAT,GLSendo,GLSmyo,GLSepi and TLSendo,TLSmyo,TLSepi were finally included for the construction of prediction models.The average AUC values of each model after 3-fold cross-validation within the training set were DT(AUC=0.758),KNN(AUC=0.865),LR(AUC=0.840),NB(AUC=0.855),RF(AUC=0.864),and XGBoost(AUC=0.786).3.The AUC,accuracy,sensitivity or recall,specificity,positive predictive value or precision,negative predictive value,and F1 scores of DT,KNN,LR,NB,RF,and XGBoost in the test set were 0.7265(95% CI: 0.5697~0.8833),0.7660,0.7059,0.8000,0.6667,0.8276,0.6857;0.8941(95% CI: 0.8018~0.9863),0.8085,0.7059,0.8667,0.7500,0.8387,0.7273;0.8412(95% CI: 0.7223~0.9601),0.7660,0.5882,0.8667,0.7143,0.7879,0.6451;0.8412(95% CI: 0.7225~0.9599),0.7872,0.5882,0.9000,0.7692,0.7941,0.6667;0.8314(95% CI: 0.7092~0.9536),0.7872,0.7059,0.8333,0.7059,0.8333,0.7059;0.7735(95% CI: 0.6443~0.9027),0.8085,0.6471,0.9000,0.7857,0.8182,0.7097.4.The calibration curves showed that the Brier scores of DT,KNN,LR,NB,RF,and XGBoost in the test set were 0.196(0.103~0.288),0.135(0.076~0.193),0.146(0.092~0.200),0.166(0.080~0.252),0.157(0.097~0.217),0.191(0.078~0.305),respectively,and the Brier scores of each model were <0.25.The prediction probabilities of the KNN model were in good agreement with the actual observed values,and the prediction probabilities of DT were in relatively poor agreement with the actual observed values.The KNN was finally screened as the best prediction model.Conclusions:The machine learning model based on the EAT thickness and myocardial longitudinal stratification strain parameters measured by echocardiography can effectively predict the severity of coronary artery lesions in CAD patients.KNN is superior to DT,LR,NB,RF,and XGBoost in determining the severity of coronary artery lesions in CAD patients and has high clinical application value. |
PDF Full Text Request