Exploring Pre-trained General-purpose Audio Representations for Heart Murmur Detection

TL;DR

The paper investigates whether general-purpose, pre-trained audio representations can transfer to heart murmur detection in settings where domain-specific data are scarce. It evaluates multiple architectures pre-trained on AudioSet, with M2D delivering state-of-the-art performance on the CirCor DigiScope task and ensembles offering further gains. The findings demonstrate that pre-training on large general audio data is valuable for heart sound analysis and that combination of representations can boost detection across classes, including the Unknown category. These results suggest practical potential for deploying transfer-learned audio representations in clinical auscultation workflows, with code made publicly available for replication and extension.

Abstract

To reduce the need for skilled clinicians in heart sound interpretation, recent studies on automating cardiac auscultation have explored deep learning approaches. However, despite the demands for large data for deep learning, the size of the heart sound datasets is limited, and no pre-trained model is available. On the contrary, many pre-trained models for general audio tasks are available as general-purpose audio representations. This study explores the potential of general-purpose audio representations pre-trained on large-scale datasets for transfer learning in heart murmur detection. Experiments on the CirCor DigiScope heart sound dataset show that the recent self-supervised learning Masked Modeling Duo (M2D) outperforms previous methods with the results of a weighted accuracy of 0.832 and an unweighted average recall of 0.713. Experiments further confirm improved performance by ensembling M2D with other models. These results demonstrate the effectiveness of general-purpose audio representation in processing heart sounds and open the way for further applications. Our code is available online which runs on a 24 GB consumer GPU at https://github.com/nttcslab/m2d/tree/master/app/circor

Figures (1)

• Figure 1: Network architecture. The pre-trained model $f_\theta$ is used as the feature extractor, and the output of the linear layer is the prediction result. The weights $\theta$ of the $f_\theta$ are initialized with pre-trained weight parameters, and all network layer parameters, including $\theta$, are trained in a fine-tuning fashion.