Graph-based feature extraction and classification of wet and dry cough signals: a machine learning approach

Area of Science:

• Medical Signal Processing
• Complex Network Theory
• Respiratory Medicine

Background:

• Cough sounds are key indicators of respiratory diseases like COVID-19.
• Differentiating between wet (WE) and dry (DE) coughs is crucial for diagnosis.
• Traditional analysis methods may not fully capture the complexities of cough acoustics.

Purpose of the Study:

• To explore graph-based features for analyzing cough signals.
• To differentiate between wet (WE) and dry (DE) coughs using complex network analysis.
• To assess the potential of these features for remote auscultation and respiratory illness detection.

Main Methods:

• Analysis of 115 cough signals (WE and DE) using spectral and complex network approaches.
• Time-domain and wavelet analysis to observe signal phases and airflow dynamics.
• Application of supervised (quadratic SVM, neural net pattern recognition - NN) and unsupervised (principal component analysis) machine learning techniques.

Main Results:

• Complex network features (degree centrality, eigenvector centrality, transitivity, graph density, graph entropy) effectively distinguished between WE and DE coughs.
• Wavelet analysis indicated respiratory tract turbulence in WE.
• Supervised MLTs, particularly NN, achieved higher classification accuracy (93.90%) and precision (97.00%) compared to unsupervised methods.

Conclusions:

• Graph-based features derived from complex network analysis offer a novel method for cough sound characterization.
• The NN classifier demonstrated high efficacy in distinguishing WE and DE coughs.
• Complex network analysis holds significant potential for developing remote auscultation tools for respiratory health monitoring.