Automated Cough Analysis with Convolutional Recurrent Neural Network

Area of Science:

• Respiratory Medicine
• Artificial Intelligence
• Signal Processing

Background:

• Chronic cough significantly impacts patient health and quality of life.
• Quantitative, real-time monitoring of cough severity is crucial for clinical practice and research.
• Existing tools for measuring spontaneous coughs in daily settings are limited.

Purpose of the Study:

• To develop and evaluate a machine learning model for automated cough sound detection and classification.
• To assess the effectiveness of Mel spectrograms as feature representations for cough analysis.
• To compare various machine learning algorithms for cough monitoring.

Main Methods:

• Utilized Mel spectrograms to capture temporal and spectral characteristics of cough sounds.
• Trained and compared machine learning algorithms including decision tree, SVM, k-NN, logistic regression, random forest, and neural networks.
• Applied the model to 300 hours of audio recordings from clinical cough challenge studies.

Main Results:

• The Convolutional Recurrent Neural Network (CRNN) approach demonstrated the highest effectiveness.
• Achieved 98% accuracy in identifying individual coughs from audio data.
• Highlighted the potential of CRNNs for analyzing complex cough patterns.

Conclusions:

• Neural network models, specifically CRNNs, show significant promise for fully automated cough monitoring.
• The developed approach offers a potential solution for quantitative cough assessment.
• Further validation is required for detecting spontaneous coughs in real-life settings, particularly in patients with refractory chronic cough.