Respiratory auscultation is a cheap way to assess lung function, but crackles and wheezes are often buried in noisy clinical recordings. A new arXiv preprint introduces RespFocus, a deep learning classifier that explicitly learns temporal evidence for these transient and sustained sounds.
The method is weakly supervised, meaning it can train on coarser labels rather than requiring precise time boundaries for every event. This could make it more practical for real-world datasets, where detailed annotation is expensive and error-prone.
The abstract does not include quantitative results or comparisons with other classifiers, so the reported contribution is the architecture and training strategy rather than proven performance gains. Further evaluation would be needed to confirm how well RespFocus generalizes across recording conditions and patient populations.