Unsupervised Heart Sound Decomposition and State Estimation with Generative Oscillation Models.

Area of Science:

• Cardiology
• Signal Processing
• Machine Learning

Background:

• Phonocardiograms (PCGs) are complex signals traditionally modeled separately for oscillatory components and state transitions.
• Existing methods like state space models and factor models capture only partial aspects of PCG dynamics.
• A unified approach is needed to comprehensively model PCG generation and characteristics.

Purpose of the Study:

• To propose a novel generative probabilistic model for phonocardiograms (PCGs).
• To simultaneously capture oscillatory factors and state transitions within cardiac cycles.
• To enable unsupervised decomposition of PCGs into cardiac state-dependent oscillations.

Main Methods:

• Combining oscillation decomposition with a state space model within a unified generative probabilistic framework.
• Developing a model that reflects the mechanism of cardiac sound generation.
• Utilizing unsupervised learning for PCG analysis.

Main Results:

• The proposed model achieved superior accuracy in state estimation compared to empirical mode decomposition.
• The model demonstrated higher accuracy in detecting S2 onsets than supervised segmentation methods, especially with varied PCG signal distributions.
• Successfully decomposed PCGs into cardiac state-dependent oscillations.

Conclusions:

• The unified generative probabilistic model offers a more comprehensive approach to PCG analysis.
• The model effectively captures the interplay between oscillations and cardiac states.
• This method advances unsupervised analysis of cardiac sounds and has potential for improved diagnostic accuracy.