Enhancing Cardiac Auscultation with Machine Learning and Deep Learning: A Survey on Heart Sound Analysis

Abstract

Cardiovascular disease kills more people than anything else. Early detection changes outcomes, and heart sounds are one of the cheapest places to look for the first signs. Phonocardiogram (PCG) recordings carry information about valve function that a trained ear can pick up, which is why auscultation has survived two centuries of newer imaging and why it is still the first test a patient meets. But listening is slow. It needs experience, and two clinicians can disagree about the same murmur. Machine learning (ML) and deep learning (DL) offer a way to automate that step and make it reproducible. This review asks what the literature has actually achieved, and what it has not. We searched IEEE Xplore, SpringerLink, PubMed, ScienceDirect and Google Scholar for work published between 2015 and 2024. Section 2 gives the search strings, the screening steps and the inclusion rules. Thirty-seven primary studies survived. Each was read in full, and its dataset, signal processing chain, model and reported performance were recorded in a comparison table that covers thirteen of them in detail. Reported accuracy in this field is high and rising. It is also, taken at face value, misleading. The studies reviewed here use different corpora, different class definitions, different segmentation strategies and different validation protocols, so 99.80% in one paper and 73.6% in another are not measuring the same thing. Standardised datasets are missing. So are agreed metrics. Paediatric recordings are thin on the ground, recording conditions vary, and almost nothing here has been tested in the clinic where it would have to work. The review pulls recent advances together, then sets the reported numbers against the methodological limits that decide how much those numbers are worth, which surveys in this area tend to skip.