Detection of coronary artery disease using a triplet network and hybrid loss function on heart sound signal

• Published in: Biomedical signal processing and control Vol. 104; p. 107601
• Main Authors: Liu, Xu, Lv, Chengcong, Cao, Linchun, Guo, Xingming
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2025
• Subjects: CAD; CWT; Heart sound; Hybrid loss function; Triplet network; Hybrid loss function; CWT; Triplet network; Heart sound; CAD
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107601

•Introduced deep metric learning for CAD diagnosis via heart sound feature distances.•Developed a hybrid loss function combining triplet and cross-entropy losses to enhance accuracy.•Identified CWT as the optimal transformation method for converting heart sound signals to 2D images. Coronary artery disease (CAD) is currently one of the most common cardiovascular diseases. Its high incidence and mortality pose a serious threat to human health and cause a huge burden on the social medical system. As a traditional, simple, non-invasive and non-intrusive detection method, auscultation has important application value in the detection of CAD. At present, most deep learning methods for detecting CAD based on heart sounds mainly rely on the deep learning model framework and have not fully utilized the distance differences of heart sound data in the mapping space under different physiological states. Therefore, in this paper, the deep metric learning method is introduced to construct a three-input Triplet Network and improve the Triplet loss function by combining the cross-entropy loss, so that the classification task is also involved in the model training to obtain a more comprehensive feature expression and improve the model performance. At the same time, this paper explores the performance of Continuous Wavelet Transform (CWT), Gramian Angular Field (GAF), Mel Frequency Cepstrum Coefficient (MFCC), Markov Transition Field (MTF)and Power Spectrogram (PS) based on heart sound signals in the proposed model. Through five-fold cross-validation, the results show that CWT is the most suitable among the five methods for CAD detection. The model with CWT as input performs better than other methods, with an accuracy rate of 93.91 %, a sensitivity of 92.73 %, a specificity of 95.00 %, an F1 score of 93.89 %, and a precision of 94.14 % at the subject-level.