Detection and classification of heart opening snaps

Studenteropgave: Speciale (inkl. HD afgangsprojekt)

  • Oliver Thomsen Damsgaard
  • Simon Bruun
4. semester , Sundhedsteknologi (cand.polyt.), Kandidat (Kandidatuddannelse)
Cardiovascular diseases (CVD) are the primary cause of death around the world, but the methods for detection still rely heavily on subjective observations during auscultation followed by, in some cases, invasive examinations. New methods based on neural network and classifiers for automatic detection of heart disorders through phonocardiography (PCG) are being tested to overcome the subjective classifications within current methods. The PCG can be used to represent the hearts state, as the mechanic nature of CVD's result in unique abnormal heart sounds. Opening snaps (OS) followed by murmurs are caused by mitral stenosis, where the changed mechanical properties of the leaflets cause a snapping sound followed by a murmur due to blood turbulence. This study examines the cause of OS without an accompanying murmur, to find if this relates to calcification in the heart. This study implements parallel Fully Convolutional Networks (FCN) coupled with Long Short-Term Memory (LSTM) neural networks followed by a support vector machine (SVM) classifier to determine if an OS is present. Three networks will operate on either a filtered signal, Mel Frequency Cepstral Coefficients (MFCC) or Discrete Wavelet Transforms (DWT), as the last mentioned features have been proved useful for sound classification. In contrary to other studies, analysis of the heart cycle will only be performed on the relevant area for the specific abnormal heart sound rather than the entire cycle. Our results show that this approach is useful with a best average accuracy of 92\% and an area under curve of 0.9288. No significant results were found for the cause of an OS without accompanied murmurs for the factors examined in this study.
Udgivelsesdato6 jun. 2019
Antal sider46
Ekstern samarbejdspartnerAcarix A/S
professor Samuel Schmidt
ID: 305244447