Unsupervised Phonocardiogram Analysis With Distribution Density Based Variational Auto-Encoders.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Cardiovascular Diagnostics

Background:

• Variational Auto-Encoders (VAEs) are state-of-the-art for Phonocardiogram (PCG) analysis.
• Existing VAE models assume normal PCG signals follow a Gaussian distribution in latent space, which may not hold true.
• This assumption can limit the performance of VAE-based PCG analysis systems.

Purpose of the Study:

• To propose unsupervised methods for PCG analysis that overcome limitations of standard VAEs.
• To improve the performance of PCG analysis by employing density estimation in the latent space.
• To investigate the representation of normal PCG signals in the latent space.

Main Methods:

• Developed two novel methods: Density-Based Variational Auto-Encoder (DBVAE) and Density-Based Auto-Encoder (DBAE).
• Both methods utilize distribution density estimation within the latent space.
• Evaluated system performance on single- and multi-domain PCG datasets.

Main Results:

• The proposed DBVAE and DBAE methods significantly outperform standard VAE-based methods on PCG analysis.
• DBAE demonstrates the ability to represent normal PCG signals following Gaussian-like models in the latent space.
• DBVAE does not enforce a Gaussian-like representation but still achieves superior performance.

Conclusions:

• Unsupervised PCG analysis can be enhanced by incorporating density estimation in the latent space.
• The proposed DBVAE and DBAE offer improved performance over traditional VAE approaches for PCG signal analysis.
• These methods provide a more flexible framework for understanding PCG signal characteristics.