Study On The Value Of Machine Learning Models Constructed Based On Clinical Data As An Aid To Diagnosis In Patients With Stable Coronary Artery Disease

Background : Because of the limited accuracy of non-invasive tests to screen for coronary artery disease(CAD),many people are incorrectly or delayed in having a coronary angiogram(CAG)performed to further confirm or rule out a diagnosis.Objectives:Based on clinical data and heart sounds,to develop and validate machine learning(ML)algorithms to effectively detect CAD.Methods:We conducted a study of patients with paroxysmal precordial pain who were evaluated by an outpatient physician for possible CAD.before each patient underwent CAG,we collected relevant clinical data and collected heart sound signals using an electronic stethoscope at nine locations in the precordial region using the clinical data we constructed multiple ML models,and using the heart sound data we trained several deep learning(DL)models to detect CAD(≥50%stenosis in at least one vessel)and constructed a model for coronary stenosis degree assessment based on clinical data.Finally,the performance of each model was tested on an independent dataset.Results:In the ML model based on clinical data,269 individuals were included in the training set and 131 in the test set,in which the XGB model performed best overall with an AUC of 0.728(95% CI:0.632-0.824),and it continued to perform best in evaluating the severity of coronary lesions.In the heart sound-based model,320 individuals were included in the training and validation sets,and 80 individuals were prospectively included in the test set.the sensitivity of VGG-16 was 80.4%,the specificity was 86.2%,and the AUC was 0.833(95% CI: 0.736-0.930).The AUC of the combined diagnostic model with VGG and PTP scores was 0.915(95% CI: 0.855-0.974)and the AUC of the combined diagnostic model of VGG and DF was 0.908(95% CI: 0.845-0.971).Conclusions:DL models based on clinical data enable non-invasive and efficient CAD screening and reduce referral rates for downstream screening.