Detection of coronary heart disease based on MFCC characteristics of heart sound

Abstract

BackgroundPhonocardiogram (PCG) has been widely used to aid in the diagnosis of cardiovascular diseases. Coronary heart disease (CHD) is one of the leading causes of death, the use of heart sounds noninvasive and convenient detection of CHD, can assist the doctor’s diagnosis and treatment, and prognosis, reducing the huge burden of disease.Objective:This study proposes a CNN + attention mechanism model that can assist in the diagnosis of CHD without relying on manual extraction of heart sound features, and initial screening of CHD in groups presenting with chest pain and other symptoms. MethodsHeart sounds were preprocessed and converted into two-dimensional characteristic MFCC. The SPC-3D model designed in this study was subjected to 5-fold cross-validation and compared with seven common deep learning models.Results:The accuracy of the SPC-3D model designed in this study was 91.4993%, the precision was 89.9039%, the F1 score was 91.9398%, the sensitivity was 94.0911%, the specificity was 88.6965%, and the results were the best. Its accuracy is 3.5% higher than the basic network MobileNetV2, and 1.2% higher than that of DenseNet, which has a deep network structure.Conclusion:Compared with other models, the sensitivity and specificity of the model designed in this study are less different, indicating that the model has strong robustness. Collecting more detailed information on the patient's condition will help to further classify, and assist in the treatment, and prognosis of CHD.