A deep neural network approach to heart murmur detection using spectrogram and peak interval features

• DOI: 10.1016/j.engappai.2024.109156
• Publication Date: 2024-08-19T09:07:56Z
• Authors: Soyul Han; Taein Kang; Jungguk Lee; Narin Kim; Hyejin Won; Yeong-Hwa Kim; Wuming Gong; Il-Youp Kwak
• Keywords: Spectrogram

Congenital heart disease affects about 1% of newborns, posing risks like failure and mortality. Developing countries often lack resources for diagnosis treatment. The George B. Moody PhysioNet Challenge 2022 aims to develop systems detecting murmurs clinical outcome events using phonocardiogram (PCG) data, offering a cost-effective method diagnosing cardiac diseases without invasive procedures. We proposed deep learning model this task achieved the 5th rank out 40 teams in both Track 1 (murmur detection) 2 (clinical prediction). This paper describes our methods additional experiments. To extract features from PCG, we employed techniques including Constant Q Transform (CQT) Mel-scaled spectrogram (Mel-spectrogram) generate two-dimensional representations frequency–time domain. Additionally, extracted Peak Interval (PI) feature, which measures distance between peaks PCG data. feature is useful because peak intervals should be shorter recordings with murmurs. also considered sequence mean PI as features. Our system, titled 'Phonocardiogram-based Heart murmur Detection Spectrogram (SpectroHeart),' employs Mel-spectrogram detect assess outcomes. believe that learning-based system has great potential automatically signals PCG.