Adaptive Metadata-Guided Supervised Contrastive Learning for Domain Adaptation on Respiratory Sound Classification

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models for respiratory sound classification (RSC) face challenges due to biased datasets from inconsistent recording and imbalanced demographics.
• This bias leads to poor performance when models are deployed in real-world scenarios.

Purpose of the Study:

• To address bias in RSC datasets by exploring metadata-guided domain adaptation techniques.
• To improve the generalizability and accuracy of RSC models across different data sources.

Main Methods:

• Investigated the impact of various metadata attributes and their combinations on RSC model performance.
• Developed an advanced method for adaptive rescaling of metadata combinations to enhance domain adaptation during training.

Main Results:

• Demonstrated a significant reduction in domain dependency for RSC models.
• Achieved an improved detection accuracy score of 84.97%, a 7.37% enhancement over the baseline model.
• Validated the effectiveness of the proposed methods on both the ICBHI dataset and a custom dataset.

Conclusions:

• Metadata-guided domain adaptation is effective in mitigating bias and improving RSC model performance.
• The proposed adaptive rescaling method offers a substantial advancement for training robust and accurate respiratory sound classification models.