Leveraging unlabeled data for lung sound classification through self-supervised contrastive learning Articles uri icon

publication date

  • February 2026

start page

  • 1

end page

  • 15

issue

  • Part A, 108477

volume

  • 112

International Standard Serial Number (ISSN)

  • 1746-8094

Electronic International Standard Serial Number (EISSN)

  • 1746-8108

abstract

  • This paper focuses on the classification of lung sounds as can be instrumental in diagnosing respiratory diseases, that are one of the leading causes of death. We address two intrinsic challenges associated with the application of deep learning models to this task. The first one is the need of large databases annotated by expert medical staff. The second one is the presence of a strong class imbalance in the datasets primarily due to the predominance of non-pathological respiratory sounds. For overcoming these issues we propose a deep convolutional network model that is trained using the contrastive self-supervised learning paradigm. This technique is able to generate useful audio data representations from unlabeled data that can be effectively transferred to the target task employing a limited amount of annotated data. We have evaluated the developed systems on the well-known ICBHI dataset that contains respiratory cycles categorized into four different classes. Results show that our approach outperforms the conventional supervised learning model when the size of available labeled data is reduced. With 40% of annotated data, self-supervision achieves a relative improvement with respect to the baseline of 12.1%, 8.6%, 16% and 66.7% in score, accuracy, and sensitivity respectively, while getting a reduction of 1.0% in specificity. Finally, to corroborate our findings, we have also assessed our system on the SPRSound database, confirming the same trends. We believe that the findings in this paper enlightens the path towards the use of unlabeled data in biomedicine alleviating the need of large annotated datasets.

subjects

  • Biology and Biomedicine
  • Telecommunications

keywords

  • respiratory sound classification; wheeze; crackle; unlabeled data; class imbalance; self-supervised learning; contrastive learning; transfer learning; data augmentation