A systematic review of machine learning approaches for phonocardiogram classification

Recent advances in machine learning (ML) have accelerated automated analysis of phonocardiogram (PCG) signals, yet prior surveys often narrow their scope (e.g. omitting segmentation, focusing only on classical ML or overlooking recent deep learning [DL] trends) and provide limited methodological transparency. We present a comprehensive, Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA)-guided synthesis of PCG classification research published predominantly between 2021 and 2025, resulting in 151 studies included in the final synthesis. The systematic search was conducted across IEEE Xplore, PubMed/MEDLINE, Scopus, SpringerLink, ScienceDirect and Google Scholar, using Boolean combinations of heart sound/PCG-related terms and ML keywords. The review maps the full pipeline - acquisition, preprocessing, segmentation, feature extraction and classification - and distinguishes feature representations as single-independent-variable (SIV) and double-independent-variable (DIV). We compare classical classifiers with modern DL and hybrid architectures, and summarize public/proprietary datasets and evaluation practices. Our analysis shows that DL - especially convolutional neural network (CNN)-based approaches-dominates recent work, with growing interest in hybrid CNN-recurrent neural networks (RNNs) models and attention/transformer architectures. Segmentation is handled either explicitly (standalone or embedded) or bypassed in end-to-end designs. Despite high benchmark performance, comparability is hindered by heterogeneous datasets, non-uniform splits and metric choices; generalization under noise, device variability and pediatric vs. adult domain shift remains a key challenge. Bridging technical progress to clinical utility requires robustness, interpretability, efficient on-device inference and privacy-preserving training. We conclude with practical recommendations for standardized evaluation protocols, interpretable modeling, multimodal fusion, edge deployment and federated/multi-center learning. By integrating methodological and clinical perspectives, this review connects state-of-the-art results with the requirements of scalable, clinically viable PCG-based screening systems.