Evaluating Transfer Learning Strategies for Lung Sound Event Classification

Abstract

Computerised lung auscultation has the potential to offer automated respiratory disease follow-up in ambulatory settings. Lung sound recordings are typically analysed using Sound Event Classification (SEC) models. However, during inference, mismatches between the training and deployment data distributions can lead to significant performance degradation. Transfer Learning (TL) techniques offer a way to mitigate this problem. In this study, we evaluate SEC performance on two in house lung sound datasets using: (a) models trained on publicly available lung sound data, and (b) those models enhanced with domain+task TL, domain TL and semi-supervised domain+task TL methods. We conclude that, for our setup, domain TL results in good classification performance when only a domain shift is present. When a task shift exists between source and target data, partially labelled target data is required to obtain good task adaptation.