Estimation of Internal Acoustics from Acoustic Signals Captured on the Body Surface
[Japanese|English]
Abstract
The heart moves almost periodically during normal conditions, transitioning through four states: S1, systole, S2, and diastole. In each state, different valves open and close, producing heart sounds through their vibrations. Therefore, based on the assumption that there are multiple vibration components derived from the physical model of the valves, and that their amplitudes change according to the states of the cardiac cycle, we constructed a probabilistic generative model to represent the mechanism of heart sound generation. Furthermore, we devised a method to estimate the states of the cardiac cycle and the parameters of each component from the observed heart sounds.
Contributions
Identifying which part of the heart is producing which sound could be an important clue for determining the presence and extent of diseases. However, there has been no non-invasive method to individually listen to the various sounds occurring inside the body. By utilizing the proposed model, it is possible to extract the vibration components that exist behind the multi-channel heart sound time series.
Application example
The four-channel heart sounds recorded from the body surface of a patient with aortic valve stenosis were decomposed into eight components. Since the proposed method uses a physical model of valve vibrations, these components are considered to originate from the various valves of the heart. Further investigation into the sound source locations, based on several assumptions, suggested that component number 8 is the main component of the sound originating from the aortic valve.
Future
By advancing the development of technology to estimate the cause and location of abnormal sounds from sounds captured on the body surface, we aim to realize an AI tele-stethoscope.
Reference
- R. Shibue, M. Nakano, T. Iwata, K. Kashino, H. Tomoike, “Unsupervised heart sound decomposition and state estimation with generative oscillation models,” EMBC, 2021.
Contact
Ryohei Shibue
Biomedical Informatics Research Group, Media Information Laboratory, NTT Communication Science Laboratories