Cough activity detection for automatic tuberculosis screening

A research team from institutions including Stellenbosch University has published a paper detailing a highly accurate AI model for detecting coughs in audio recordings, a critical step for automated tuberculosis (TB) screening. The system fine-tunes Meta's large, pre-trained audio model XLS-R on a dataset of coughs from symptomatic TB patients at community care centers in South Africa and Uganda. The model's key achievement is pinpointing the exact start and end points of coughs with an average precision of 0.96 and an area under the ROC curve (AUC) of 0.99 on the test set, demonstrating near-perfect discrimination.

The researchers made a crucial optimization for real-world deployment: they found that using only the first three layers of the XLS-R network delivered the best performance while drastically reducing computational and memory needs. This streamlined version outperformed an Audio Spectrogram Transformer (AST) by 9% and a logistic regression baseline by 27% in average precision. Furthermore, a downstream TB classifier trained on coughs isolated by this AI model performed nearly as well as one trained on manually verified 'ground truth' coughs. This proves the system's practical utility, making it feasible to integrate into smartphone-based screening applications for scalable, low-cost health monitoring in resource-limited settings.