Automatic cough classification for tuberculosis screening in a real-world environment.

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Pulmonology

Background:

• Tuberculosis (TB) diagnosis can be challenging, especially in resource-limited settings.
• Accurate and rapid screening tools are crucial for effective TB control.
• Current diagnostic methods may be invasive, costly, or time-consuming.

Purpose of the Study:

• To develop and evaluate an automated system for discriminating between tuberculosis (TB) cough sounds and those from other lung ailments.
• To assess the performance of machine learning classifiers in analyzing cough audio for TB detection.

Main Methods:

• A dataset of 1358 forced cough recordings from TB patients and controls was utilized.
• Five machine learning classifiers were trained and evaluated, including logistic regression (LR) and convolutional neural networks.
• Feature selection techniques were employed to optimize classifier performance.

Main Results:

• Logistic regression, combined with sequential forward feature selection, achieved the highest performance.
• The best LR model attained an Area Under the ROC Curve (AUC) of 0.94.
• This system demonstrated 93% sensitivity and 95% specificity, exceeding WHO minimal requirements for TB triage.

Conclusions:

• Automated cough sound analysis shows significant promise as a low-cost, deployable frontline screening tool for TB.
• The developed system meets WHO triage specifications, identifying patients for further molecular testing.
• This technology can greatly benefit developing countries with a high TB burden.