A self-Supervised ConvNeXt (SCONE) framework for robust and interpretable heart sound classification

Read full text Summarise with VIRAT
Authors: M. Gayathri, Jogendra Kumar, Chandu Jagan Sekhar Madala, Shanmuk Srinivas Amiripalli, Sai Prasad Potharaju, M. V. V. Prasad Kantipudi
Cited by: 0
References: 30

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Problem Considered Accurate and reliable classification of heart sounds is critical for early screening of cardiovascular abnormalities. However, most existing deep learning approaches rely on fully supervised training and are evaluated primarily under in-distribution settings, limiting their robustness and generalizability in real-world clinical scenarios where data distributions and acquisition conditions vary. In addition, the scarcity of labelled phonocardiogram (PCG) data and the lack of systematic robustness evaluation further hinder the practical deployment of automated heart sound analysis systems. Method This study proposes a self-supervised ConvNeXt-based framework for heart sound classification that leverages contrastive representation learning to exploit unlabeled PCG recordings prior to task-specific fine-tuning. Log-Mel spectrograms are used as input representations, and the encoder is pretrained using a contrastive objective to learn invariant acoustic features, followed by supervised optimization with a lightweight classification head. To emphasize practical reliability, the model is evaluated using multi-seed experiments, validation-driven threshold calibration, and domain-aware testing protocols. In addition, Grad-CAM is employed to provide visual explanations of model predictions. Results Experiments conducted on the PhysioNet heart sound dataset demonstrate that the proposed framework achieves stable and consistent performance across multiple runs, with a favorable balance between sensitivity and specificity and a mean accuracy of approximately 94%. The results show that self-supervised pretraining improves representation robustness and reduces sensitivity to initialization and data partitioning. Visual attribution maps further indicate that the model focuses on clinically meaningful time–frequency regions associated with cardiac events, supporting the interpretability and clinical relevance of the proposed approach.

Fields of Study & Subjects

Phonocardiography and Auscultation Techniques Artificial Intelligence in Healthcare ECG Monitoring and Analysis

Publication details

Journal / source
Intelligence-Based Medicine
Volume
14
Pages
100377–100377
Publication date
25 March 2026
Language
EN