Nonnegative matrix factorization of phonocardiograms for heart rate detection

Phonocardiograms (PCGs) are recordings of the sounds and murmurs made by the heart detected through specialized microphones placed on a patient's thorax. Alongside electrocardiograms (ECGs), they are a tool used in a medical environment to assess patients' conditions relative to their cardiac rhythm. Unlike the latter, in which, during each cardiac cycle, only one main peak can be detected within the voltage-over-time graph (the so called R wave), in PCGs two distinct peaks can be observed. These peaks are associated to the first and second heart sound (S1 and S2), generated by the closure of specific valves within the heart. In the following we shall refer to R waves and S1, S2 sounds as 'cardiac events'. In order to extrapolate the heart's activity from ECGs or PCGs, one needs to detect all cardiac events within the signal of choice. When it comes to ECGs, this process is relatively straightforward since only one R wave ought to be identified during each cycle. Moreover, such signals usually contain very low levels of noise, mainly caused by powerline interference, which can be easily removed using notch filters. On the other hand, event detection within PCGs is a quite challenging task. Indeed, not only do we need to detect two sounds each cycle, but also the signal itself is often severely contaminated by many different types of noise, such as the patient's movement, ambient sources, microphone movement or other body-related murmurs. As a consequence, the analysis of PCGs is often carried out with the aid of a synchronous ECG signal and requires a careful denoising of the audio file through digital filtering and signal envelope estimation. The objective of the dissertation was to develop a method of detecting cardiac events within PCG signals that does not rely on the knowledge of ECGs. In particular, we achieved our goal by leveraging the modelling and learning capabilities of Nonnegative Matrix Factorization (NMF) applied to the spectrogram of PCGs.