A Comparison of Solicited and Longitudinal Cough Sounds for Tuberculosis Detection

Tuberculosis (TB) caused an estimated 1.25 million deaths in 2023, exceeding the number affected by COVID-19. Existing diagnostic tools are costly and inaccessible in lowresource settings, leading to delays and ongoing transmission. A rapid and affordable screening method is urgently needed to identify and refer suspected TB cases for confirmation. One promising approach is deep learning using cough sounds, which offers a low-cost, scalable solution. However, building an effective deep learning model depends heavily on data, particularly the balance between data quality and quantity. Two types of datasets were evaluated: solicited (supervised recording) and longitudinal (unsupervised recording). Results show that supervised recordings achieved higher performance than unsupervised ones when using the same data size (79 % vs. 71 % of accuracy). However, increasing the amount of unsupervised data significantly improved performance, reaching 91% accuracy-highlighting the benefit of larger datasets. Interestingly, combining solicited and longitudinal data did not enhance performance, likely due to the small proportion of supervised data.