Similarity-Based Signal Embedding for Heart Sound Abnormality Detection

Heart sound (HS) analysis is important for early detection of cardiovascular abnormalities. Although deep learning methods have shown strong performance, they often require large datasets and high-dimensional inputs. In this study, a similarity-based framework is proposed for heart sound abnormality detection using low-dimensional statistical features. Experiments were conducted on the publicly available HLS-CMDS dataset, and only heart sound recordings were used. Each signal was divided into five segments, and time- and frequency-domain statistical features were extracted. Instead of direct classification, a pairwise similarity learning approach was adopted. Feature-difference vectors were constructed between segment pairs, and an RBF-kernel SVM was trained to learn similarity relationships. The learned similarity function was then used to generate a prototype-based embedding for each signal. The proposed method provides a computationally efficient and interpretable alternative to deep learning models, especially for small-scale biomedical datasets.