Automated Wheezing Segmentation with Task-Specific Weighted Loss
Abstract Accurate segmentation of wheezing events may support clinical decision-making and patient monitoring. Previous studies have largely focused on classifying abnormal respiratory sounds at the recording or respiratory-cycle level, whereas precise temporal localization of wheezing onset and offset remains challenging. We propose Wheezing Forecasting & Audio Segmentation Technology (W-FAST), an architecture for precise temporal segmentation of wheezing events. To address the severe class imbalance in clinical respiratory sound data, W-FAST employs a task-specific weighted loss incorporating a weighting factor α. W-FAST achieved a mean ICBHI score of 0.673 ± 0.017 on the ICBHI 2017 dataset. When the same ICBHI scoring scheme was applied to the SNUBH dataset, W-FAST achieved a mean score of 0.946 ± 0.016. In the SNUBH dataset, the ratio of wheezing to non-wheezing blocks was approximately 1:55 among valid 0.32-s blocks after zero-padded regions were masked. The proposed loss helped stabilize training under this severe imbalance and improved segmentation performance. These results suggest that W-FAST may provide an effective framework for fine-grained temporal localization of wheezing in clinical respiratory sound recordings.
Authors
- June-Woo Kim (ORCID: https://orcid.org/0000-0003-0111-300X)
- Kyunghoon Kim (ORCID: https://orcid.org/0000-0002-0707-6839)
- Hyunwoo Kim (ORCID: https://orcid.org/0009-0002-1382-5848)
- Wonwoo Shin
- Daehwan Hwang
Institutions
- Seoul National University Bundang Hospital (KR)
- Wonkwang University (KR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-10978827/v1
- Primary Topic
- Phonocardiography and Auscultation Techniques
- Type
- preprint