Can AI identify tuberculosis from cough audio recordings with better accuracy than human clinicians ?

Tuberculosis remains a leading infectious killer, where early diagnosis could dramatically improve outcomes. Could artificial intelligence outperform trained clinicians in identifying tuberculosis from cough sounds alone? The next section explores the science behind this emerging diagnostic approach.

Background

Tuberculosis (TB) is a leading infectious cause of death globally, with early diagnosis critical for successful treatment. Cough acoustics contain unique biomarkers that may reflect underlying pulmonary pathology, including TB-specific signatures. AI models—particularly convolutional neural networks leveraging transfer learning—have been trained on crowdsourced cough datasets to detect TB with reported sensitivities and specificities of approximately 90–95%. Such systems aim to enable remote, low-cost screening in resource-limited settings, addressing gaps where access to clinical expertise or laboratory diagnostics is constrained. However, performance heavily relies on high-quality audio recordings; real-world deployment faces challenges from ambient noise, variability in recording equipment, and overlapping respiratory conditions. Current validation remains largely dataset-dependent, and broader clinical implementation awaits real-world trials and regulatory clearance. WHO emphasizes that rigorous validation across diverse populations is essential to ensure equitable and reliable diagnostic performance.