Enhancing Respiratory Sound Classification Based on Open-Set Semi-Supervised Learning

The classification of respiratory sounds is crucial in diagnosing and monitoring respiratory diseases. However, auscultation is highly subjective, making it challenging to analyze respiratory sounds accurately. Although deep learning has been increasingly applied to this task, most existing approaches have primarily relied on supervised learning. Since supervised learning requires large amounts of labeled data, recent studies have explored self-supervised and semi-supervised methods to overcome this limitation. However, these approaches have largely assumed a closed-set setting, where the classes present in the unlabeled data are considered identical to those in the labeled data. In contrast, this study explores an open-set semi-supervised learning setting, where the unlabeled data may contain additional, unknown classes. To address this challenge, a distance-based prototype network is employed to classify respiratory sounds in an open-set setting. In the first stage, the prototype network is trained using labeled and unlabeled data to derive prototype representations of known classes. In the second stage, distances between unlabeled data and known class prototypes are computed, and samples exceeding an adaptive threshold are identified as unknown. A new prototype is then calculated for this unknown class. In the final stage, semi-supervised learning is employed to classify labeled and unlabeled data into known and unknown classes. Compared to conventional closed-set semi-supervised learning approaches, the proposed method achieved an average classification accuracy improvement of 2%–5%. Additionally, in cases of data scarcity, utilizing unlabeled data further improved classification performance by 6%–8%. The findings of this study are expected to significantly enhance respiratory sound classification performance in practical clinical settings.