A systematic review of machine learning approaches for phonocardiogram classification

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Automated phonocardiogram (PCG) analysis using machine learning (ML) is rapidly advancing.
• Previous reviews often have limited scope, omitting key areas like segmentation or recent deep learning (DL) trends.
• Methodological transparency in PCG classification research needs improvement.

Purpose of the Study:

• To provide a comprehensive, PRISMA-guided synthesis of PCG classification research from 2021-2025.
• To map the entire PCG analysis pipeline, from acquisition to classification.
• To compare classical ML with modern DL and hybrid architectures.

Main Methods:

• Systematic literature search across major scientific databases (IEEE Xplore, PubMed, Scopus, etc.).
• Inclusion of 151 studies focusing on PCG classification and ML.
• Analysis of segmentation techniques, feature representations (SIV, DIV), and classifier types (classical, DL, hybrid).

Main Results:

• Deep learning (DL), especially Convolutional Neural Networks (CNNs), dominates recent PCG research.
• Growing interest in hybrid CNN-RNN models and attention/transformer architectures.
• Challenges persist in dataset heterogeneity, evaluation metrics, and generalization across different conditions (noise, devices, age groups).

Conclusions:

• DL approaches show high benchmark performance but require standardization for clinical utility.
• Future work should focus on robustness, interpretability, efficient on-device inference, and privacy-preserving training.
• Recommendations include standardized evaluation, interpretable models, multimodal fusion, and federated learning for scalable PCG screening.