A cough detection method based on the conformer-BiLSTM model.

Area of Science:

• Computational linguistics
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Cough detection is crucial for respiratory disease diagnosis.
• Manual cough sound segmentation is time-consuming and subjective.
• Real-world cough recordings present detection challenges due to environmental and device variability.

Purpose of the Study:

• To develop an automated and accurate method for cough sound event detection.
• To improve the efficiency and objectivity of cough analysis.
• To address the difficulties in detecting coughs from diverse real-world audio data.

Main Methods:

• A hybrid model combining conformer and bidirectional long short-term memory (BiLSTM) networks was proposed.
• The BiLSTM model captured contextual information, while the conformer model extracted global features using multi-head attention residual connections.
• The model was trained and validated on the publicly available Vocalsound dataset, comprising 500 audio files with 1928 annotated cough sounds.

Main Results:

• The hybrid model achieved 97.22% sensitivity and 97.08% specificity on the annotated dataset.
• In a clinical dataset evaluation, the model demonstrated 98.45% sensitivity and 92.7% specificity.
• An average time overlap rate of 81.12% was recorded in the clinical dataset, indicating precise event localization.

Conclusions:

• The proposed conformer-BiLSTM hybrid model offers a robust and accurate solution for automated cough sound detection.
• This approach significantly enhances the efficiency and reliability of cough analysis in respiratory disease research.
• The model's high performance in both public and clinical datasets suggests its potential for real-world applications in healthcare.