Paediatric Respiratory Sound Classification by Integrating Local Feature Extraction and Global Context Modeling

Proposes a novel architecture that integrates local feature extraction with global context modeling for paediatric respiratory sound classification. Prior respiratory sound classification research is extensive for adults, whereas work on paediatric respiratory sounds is still scarce, and paediatric sounds typically contain more high-frequency components, posing unique challenges. Presented as an architectural design in the abstract, constituting a conceptual and design-level contribution without ablation or comparison details in the abstract.

Designs the Mel_Grouper module as a front-end to enhance local pathological representations. Targets the inherent difference that paediatric respiratory sounds contain more high-frequency components by adding a dedicated front-end module to strengthen local pathological features. A component-level design description whose standalone effect depends on experiments reported in the full text.

Feeds the Mel_Grouper output into the Transformer-based Mel_Encoder to fuse global context. Introduces global context modeling beyond local features, forming a combined local-and-global processing path. A component-level design description; the abstract does not report separate quantitative results for this module.

Achieves state-of-the-art performance on the four SPRSound subtasks and further experiments on a real-world paediatric respiratory sound dataset. Improves over the previous best results by 3.37%, 3.06%, 6.83% and 5.36% respectively, and adds experiments under real clinical conditions. Improvement margins are stated quantitatively as percentages in the abstract; specific values for the real-world dataset experiments are not listed in the abstract.

The results address paediatric respiratory sound classification, applying to the paediatric respiratory sound setting represented by the SPRSound dataset and to the real-world paediatric respiratory sound dataset described in the paper; the method centres on combining a Mel_Grouper front-end with a Transformer encoder, designed for the higher high-frequency content of paediatric sounds. The paper states that code is available, supporting reproduction and extension under the same task setting.

Specific performance values on the real-world paediatric respiratory sound dataset, the independent contribution of each module, and stability across different data splits or clinical conditions are not elaborated in the abstract; the code link in the abstract is also not given as a complete address. These are directions a reader may wish to watch when reading further.