Enhancing cross-domain robustness in phonocardiogram signal classification using domain-invariant preprocessing and...

Area of Science:

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

Background:

• Phonocardiogram (PCG) signal analysis offers a non-invasive method for diagnosing cardiovascular diseases.
• Current machine learning (ML) approaches for PCG analysis struggle with performance variations across different datasets due to varying data acquisition settings.
• This variability significantly impacts the reliability of automated disease detection systems.

Purpose of the Study:

• To investigate the impact of data acquisition parameter variations on PCG data from different databases.
• To develop robust methods for PCG-based cardiovascular disease detection that are resilient to cross-dataset variations.
• To enhance the real-world applicability of automated cardiac screening systems.

Main Methods:

• Employed a combination of domain-invariant preprocessing, transfer learning, and domain-balanced variable hop fragment selection (DBVHFS).
• Domain-invariant preprocessing normalized PCG signals to minimize stethoscope and environmental variations.
• Transfer learning utilized pre-trained audio models for generalized feature representation, and DBVHFS ensured balanced training fragment distribution across all domains.

Main Results:

• The proposed method was evaluated on six independent PhysioNet/CinC Challenge 2016 PCG databases using a leave-one-dataset-out cross-validation strategy.
• Achieved a relative improvement of 5.92% in unweighted average recall and 17.71% in sensitivity compared to existing methods.
• Demonstrated superior performance in cross-dataset evaluations, highlighting the system's robustness.

Conclusions:

• The developed methods effectively address variations in PCG data from diverse sources.
• The proposed approach shows significant potential for improving the reliability and implementation of automated cardiac screening systems in clinical practice.
• Enhanced robustness against data variability paves the way for more widespread adoption of AI in cardiovascular diagnostics.