Tuberculosis Detection from Cough Recordings Using Bag-of-Words Classifiers

Area of Science:

• Medical Informatics
• Machine Learning
• Respiratory Medicine

Background:

• Tuberculosis (TB) detection remains a global health challenge, necessitating innovative diagnostic approaches.
• Cough recordings offer a non-invasive and accessible data source for potential TB identification.

Purpose of the Study:

• To evaluate the efficacy of Bag-of-Words (BoW) classifiers for reliable tuberculosis detection using cough recordings.
• To assess the impact of different feature extraction and encoding strategies on classifier performance.

Main Methods:

• Utilized Bag-of-Words classifiers with distinct feature extraction and encoding strategies.
• Conducted experiments on two large, independent datasets, including augmented recordings.
• Employed repeated k-fold cross-validation and external datasets for performance assessment.

Main Results:

• Achieved up to 0.77 accuracy and 0.84 Area Under Curve (AUC).
• Demonstrated favorable performance compared to existing tuberculosis detection methods.
• Showcased robustness across various parameter combinations and dataset augmentations.

Conclusions:

• Bag-of-Words classifiers provide a reliable and effective method for tuberculosis detection from cough sounds.
• The proposed approach offers a promising, non-invasive tool for augmenting TB diagnostics.
• Further research can explore advanced feature engineering and deep learning integration.