COVID-19 Detection from Cough Recordings Using Bag-of-Words Classifiers

Area of Science:

• Computational biology
• Medical informatics
• Signal processing

Background:

• Accurate COVID-19 detection is crucial for public health.
• Cough recordings offer a non-invasive method for disease detection.
• Machine learning approaches are increasingly used for analyzing biomedical signals.

Purpose of the Study:

• To evaluate the reliability of detecting COVID-19 using cough recordings and bag-of-words classifiers.
• To assess the impact of different feature extraction and encoding strategies on detection performance.
• To compare machine learning approaches with Convolutional Neural Networks (CNNs) for cough-based COVID-19 detection.

Main Methods:

• Utilized bag-of-words classifiers for cough sound analysis.
• Evaluated four distinct feature extraction procedures and four encoding strategies.
• Assessed input and output fusion approaches.
• Performed comparative analysis against 2D CNNs.
• Conducted experiments on COUGHVID and COVID-19 Sounds datasets.

Main Results:

• Sparse encoding demonstrated superior performance across evaluated metrics (AUC, accuracy, sensitivity, F1-score).
• The chosen encoding strategy showed robustness irrespective of feature type or codebook dimension.
• Fusion approaches were investigated to potentially enhance detection accuracy.
• Bag-of-words classifiers, particularly with sparse encoding, proved effective for COVID-19 detection from coughs.

Conclusions:

• Sparse encoding is a highly effective strategy for COVID-19 detection from cough sounds using bag-of-words models.
• The findings suggest that cough analysis using machine learning can be a reliable diagnostic tool.
• Further research can explore optimizing these methods for real-world clinical applications.