Detection of tuberculosis by automatic cough sound analysis

Area of Science:

• Medical Diagnostics
• Bioacoustics
• Computational Biology

Background:

• Tuberculosis (TB) is a leading global cause of death.
• Current diagnostic methods can be expensive and inaccessible in low-income countries.
• Cough is a primary symptom used for TB screening, but self-reporting lacks accuracy.

Purpose of the Study:

• To investigate an automated method for tuberculosis screening using cough sound analysis.
• To develop and validate statistical classifiers for distinguishing TB-positive coughs from healthy coughs.

Main Methods:

• Collected a database of cough audio recordings.
• Developed statistical classifiers utilizing short-term spectral information.
• Evaluated classifier performance using accuracy and Area Under the Curve (AUC).

Main Results:

• Classifiers achieved 78% accuracy and 0.95 AUC in distinguishing TB coughs.
• Accuracy improved to 82% when combined with five clinical measurements.
• A sensitivity of 95% at 72% specificity was achievable with threshold adjustment.

Conclusions:

• Automated cough sound classification is a potential low-cost, low-complexity screening method for TB.
• The system utilizes spectral information beyond human auditory perception.
• This approach could significantly aid TB detection in resource-limited areas.