[Paper Review] An Update on a Progressively Expanded Database for Automated Lung Sound Analysis

Fu-Shun Hsu, Shang-Ran Huang+6 more|arXiv (Cornell University)|Feb 8, 2021

Phonocardiography and Auscultation Techniques12 references4 citations

TL;DR

This paper presents HF_Lung_V2, an expanded open-access lung sound database with a 1.45x increase in audio files over HF_Lung_V1, enabling improved deep learning models for detecting inhalation, exhalation, continuous (CAS), and discontinuous adventitious sounds (DAS). Despite improved performance in inhalation, exhalation, and CAS detection, DAS detection remains limited by label quality and sound overlapping, highlighting the need for refined labeling and overlapping mitigation strategies.

ABSTRACT

Purpose: We previously established an open-access lung sound database, HF_Lung_V1, and developed deep learning models for inhalation, exhalation, continuous adventitious sound (CAS), and discontinuous adventitious sound (DAS) detection. The amount of data used for training contributes to model accuracy. Herein, we collected larger quantities of data to further improve model performance. Moreover, the issues of noisy labels and sound overlapping were explored. Methods: HF_Lung_V1 was expanded to HF_Lung_V2 with a 1.45x increase in the number of audio files. Convolutional neural network-bidirectional gated recurrent unit network models were trained separately using the HF_Lung_V1 (V1_Train) and HF_Lung_V2 (V2_Train) training sets and then tested using the HF_Lung_V1 (V1_Test) and HF_Lung_V2 (V2_Test) test sets, respectively. Segment and event detection performance was evaluated using the F1 scores. Label quality was assessed. Moreover, the overlap ratios between inhalation, exhalation, CAS, and DAS labels were computed. Results: The model trained using V2_Train exhibited improved F1 scores in inhalation, exhalation, and CAS detection on both V1_Test and V2_Test but not in DAS detection. Poor CAS detection was attributed to the quality of CAS labels. DAS detection was strongly influenced by the overlapping of DAS labels with inhalation and exhalation labels. Conclusion: Collecting greater quantities of lung sound data is vital for developing more accurate lung sound analysis models. To build real ground-truth labels, the labels must be reworked; this process is ongoing. Furthermore, a method for addressing the sound overlapping problem in DAS detection must be formulated.

Motivation & Objective

• To enhance the accuracy of automated lung sound analysis by expanding the HF_Lung_V1 database into HF_Lung_V2 with significantly more audio data.
• To evaluate the impact of increased training data on deep learning model performance for lung sound segmentation and event detection.
• To identify and address persistent challenges in label quality and sound overlapping, particularly affecting DAS detection.
• To provide a foundation for future research by releasing a progressively growing, open-access lung sound database with improved data quality and scalability.

Proposed method

• The HF_Lung_V1 database was expanded to HF_Lung_V2, increasing the number of audio files by 1.45x through systematic data collection.
• Convolutional neural network–bidirectional gated recurrent unit (CNN-BGRU) models were trained separately on HF_Lung_V1 (V1_Train) and HF_Lung_V2 (V2_Train) datasets.
• Model performance was evaluated using F1 scores on two test sets: HF_Lung_V1 (V1_Test) and HF_Lung_V2 (V2_Test) to assess generalization and improvement.
• Label quality was assessed to identify noisy labels, particularly affecting CAS detection performance.
• Overlap ratios between inhalation, exhalation, CAS, and DAS labels were computed to quantify temporal misalignment issues.
• The study employed a comparative evaluation framework to isolate the effects of data quantity versus label quality and overlapping on model performance.

Experimental results

Research questions

• RQ1 How does increasing the size of the lung sound database impact the performance of deep learning models in detecting inhalation, exhalation, CAS, and DAS?
• RQ2 To what extent do label quality issues, particularly in CAS annotations, affect model detection accuracy?
• RQ3 How does the overlapping of DAS labels with inhalation and exhalation labels influence DAS detection performance?
• RQ4 Can a larger training dataset compensate for poor label quality or overlapping in lung sound analysis?
• RQ5 What are the key bottlenecks in current automated lung sound analysis, and how can they be addressed through data and model improvements?

Key findings

• The model trained on HF_Lung_V2 (V2_Train) achieved higher F1 scores than the V1_Train model on both V1_Test and V2_Test sets for inhalation, exhalation, and CAS detection.
• DAS detection performance did not improve with the expanded dataset, indicating that data quantity alone is insufficient to overcome underlying data quality issues.
• Poor CAS detection was directly linked to low-quality or noisy labels in the CAS annotation, which limited model performance despite increased training data.
• DAS detection was significantly hampered by overlapping labels with inhalation and exhalation, suggesting temporal misalignment as a major confounding factor.
• The study confirms that data quantity improves performance only when label quality and temporal alignment are addressed, highlighting the need for reworking ground-truth labels.
• The authors conclude that future progress requires not only larger datasets but also improved labeling protocols and methods to resolve sound overlapping in DAS detection.