HEART SOUND CLASSIFICATION VIA VISION TRANSFORMER MODELS

Abstract

The automatic heart sound classification is an integral part of the early diagnosis of cardiovascular diseases(CVDs). Even though advances in medical technologies allow us to diagnose many CVDs, it remains one of the leading causes of death worldwide due to its absence of symptoms at the initial stages. Thus, there is a huge demand to develop other methods of identifying heart sound abnormalities that are less expensive, simple, and applicable. Several audio feature extraction methods, in combination with classification models, have been developed over time. However, existing feature extraction methods are sensitive to noise, which negatively impacts the performance of the heart sound classification model. In addition, there is a strong need to develop models more sensitive to heart sound abnormalities in patients. In this work, we address the limitations of extracted features by using spectrogram images that are taken from Discrete Fourier Transform, and introducing them to Vision Transformer Model. Results of our experiments on the benchmark of PhysioNet Heart Sound Dataset show that the proposed method outperforms existing methodologies with an accuracy of 0.925 and with a sensitivity score of 0.955