A data-augmentation technique for respiratory sound classification that improves a model's ability to generalize across datasets recorded under different conditions. Inspired by Mixup, Lungmix blends raw respiratory waveforms using loudness-derived and random masks and interpolates the corresponding labels by their semantic meaning, producing plausible new training examples. Students learn how augmenting in the waveform domain with semantically interpolated multi-labels helps a classifier learn representations that transfer to unseen respiratory-sound datasets.
Respiratory sound classification plays a pivotal role in diagnosing respiratory diseases. While deep learning models have shown success with various respiratory sound datasets, our experiments indica…