[Paper Review] Semi-supervised and Unsupervised Methods for Heart Sounds Classification in Restricted Data Environments

Balagopal Unnikrishnan, Pranshu Ranjan Singh+2 more|arXiv (Cornell University)|Jun 3, 2020

Phonocardiography and Auscultation Techniques14 citations

TL;DR

This paper proposes semi-supervised and unsupervised methods for heart sounds classification in low-resource data environments using the PhysioNet/CinC 2016 dataset. It introduces a GAN-based semi-supervised approach that improves AUROC over supervised baselines with limited labels, and demonstrates that 1D CNN autoencoder features combined with one-class SVM achieve strong anomaly detection performance (AUROC up to 0.842) without any labeling, enabling efficient data-efficient diagnosis and dataset curation.

ABSTRACT

Automated heart sounds classification is a much-required diagnostic tool in the view of increasing incidences of heart related diseases worldwide. In this study, we conduct a comprehensive study of heart sounds classification by using various supervised, semi-supervised and unsupervised approaches on the PhysioNet/CinC 2016 Challenge dataset. Supervised approaches, including deep learning and machine learning methods, require large amounts of labelled data to train the models, which are challenging to obtain in most practical scenarios. In view of the need to reduce the labelling burden for clinical practices, where human labelling is both expensive and time-consuming, semi-supervised or even unsupervised approaches in restricted data setting are desirable. A GAN based semi-supervised method is therefore proposed, which allows the usage of unlabelled data samples to boost the learning of data distribution. It achieves a better performance in terms of AUROC over the supervised baseline when limited data samples exist. Furthermore, several unsupervised methods are explored as an alternative approach by considering the given problem as an anomaly detection scenario. In particular, the unsupervised feature extraction using 1D CNN Autoencoder coupled with one-class SVM obtains good performance without any data labelling. The potential of the proposed semi-supervised and unsupervised methods may lead to a workflow tool in the future for the creation of higher quality datasets.

Motivation & Objective

• Address the challenge of limited labeled heart sound data in clinical settings, where labeling is costly and time-consuming.
• Explore semi-supervised and unsupervised learning as viable alternatives to supervised methods under data scarcity.
• Evaluate the performance of GAN-based semi-supervised learning and autoencoder-based anomaly detection for heart sound classification.
• Investigate the robustness of unsupervised methods under data contamination and label scarcity.
• Develop a framework for creating higher-quality datasets using low-labeled or no-labeled data approaches.

Proposed method

• Utilizes the PhysioNet/CinC 2016 Challenge dataset containing 3,240 labeled heart sound recordings (normal and abnormal) for evaluation.
• Applies a GAN-based semi-supervised learning method that leverages unlabelled data to improve model generalization and AUROC performance.
• Employs a 1D Convolutional Neural Network (CNN) autoencoder for unsupervised feature learning, extracting latent representations from heart sound signals.
• Uses one-class SVM and Isolation Forest for anomaly detection, with features derived from autoencoder embeddings or reconstruction loss.
• Trains autoencoders on both pure normal data and mixed data to assess robustness to data contamination.
• Evaluates performance using AUROC as the primary metric across all methods.

Experimental results

Research questions

• RQ1 Can semi-supervised learning with a GAN-based approach improve AUROC performance compared to supervised baselines when only limited labeled data is available?
• RQ2 How effective is unsupervised feature learning via a 1D CNN autoencoder combined with one-class SVM for detecting abnormal heart sounds without any labeled data?
• RQ3 Does the choice of training data (pure normal vs. mixed normal and abnormal) significantly affect the performance of unsupervised anomaly detection in heart sound classification?
• RQ4 How do different anomaly detection methods (e.g., one-class SVM vs. Isolation Forest) compare when using autoencoder embeddings or reconstruction loss as input features?
• RQ5 Can unsupervised methods identify ambiguous or borderline samples that are difficult to classify, enabling better data selection for active learning or pseudo-labeling?

Key findings

• The GAN-based semi-supervised method achieves higher AUROC than the supervised baseline across all label proportions, demonstrating superior performance with limited labeled data.
• The unsupervised method using 1D CNN autoencoder features and one-class SVM achieves an AUROC of 0.842 when trained on only normal samples, outperforming Isolation Forest and other baselines.
• Autoencoder embeddings yield better anomaly detection performance than reconstruction loss, especially in low-label and contamination scenarios.
• Even with data contamination, the autoencoder trained on normal data alone produces robust latent representations that generalize well to abnormal samples.
• The unsupervised approach with autoencoder and one-class SVM achieves AUROC of 0.828 when trained on mixed data, indicating strong generalization despite contamination.
• The proposed methods show strong potential for use in active learning and pseudo-labeling pipelines, particularly for selecting ambiguous or borderline cases for human review.