A Contrastive Embedding-Based Domain Adaptation Method for Lung Sound Recognition in Children Community-Acquired Pneumonia

Lung sound analysis has been used for assessing the lung conditions of children with community-acquired pneumonia (CAP). However, the inconsistent data distribution, caused by variant CAP symptoms appearing at different ages of children, limits the generalization ability of most diagnostic models. The data scarcity will further exacerbate this problem. Therefore, we propose a contrastive embedding-based domain adaptation network (CEDANN) to eliminate individual differences and alleviate data scarcity for improving the generalization ability. It forces the embedding layer to learn the subject-independent but task-dependent features by the adversarial learning between the domain classifier and the task classifier. Multiple contrast input tuples are constructed by combining samples from different classes to increase input combinations to alleviate data scarcity. The proposed method is evaluated on a multi-center clinical dataset. The subject-independent experiments show that CEDANN improves the sensitivity from 45.93% to 59.06% and the specificity from 35.07% to 59.55% in identifying CAP-confirmed, symptomatic relief, and recovery children. It demonstrates the effectiveness of CEDANN in the diagnosis and prognosis of children CAP.