Automatic cough detection from realistic audio recordings using C-BiLSTM with boundary regression

Area of Science:

• Medical technology
• Artificial intelligence
• Signal processing

Background:

• Accurate automatic cough detection is crucial for diagnosing and monitoring respiratory diseases, including COVID-19.
• Existing detection methods often fall short of practical requirements for real-world audio recordings.
• The integrity of cough events is vital for subsequent clinical analysis.

Purpose of the Study:

• To develop a novel deep learning model for accurate cough detection and precise localization of cough events in patient audio recordings.
• To introduce a new benchmark dataset for cough detection research.
• To enhance the performance of automatic cough detection systems.

Main Methods:

• A deep convolutional bidirectional long short-term memory (C-BiLSTM) model was developed, incorporating convolutional layers to improve feature extraction and temporal information preservation.
• Boundary regression was embedded into the model to achieve more accurate localization of cough event boundaries.
• A realistic audio dataset, the Corp Dataset, was created, containing 168 hours of recordings with 9969 coughs from 42 patients.

Main Results:

• The C-BiLSTM model with boundary regression achieved a sensitivity of 84.13%, a specificity of 99.82%, and an intersection-over-union (IoU) of 0.89.
• The proposed system significantly outperformed existing related models across all cough detection criteria.
• The Corp Dataset, containing extensive labeled cough data, is publicly available for research.

Conclusions:

• The developed C-BiLSTM model with boundary regression represents a state-of-the-art system for cough detection in realistic audio data.
• The model's ability to accurately classify and locate cough events lays a foundation for improved respiratory disease monitoring.
• The open-source Corp Dataset serves as a valuable resource and benchmark for advancing cough detection research.