Vibrome biosignatures technology decodes the “music of the body” and is able to identify and visualize lung abnormalities with high sensitivity and moderate specificity. This innovation has the potential to support TB screening and diagnosis without radiation, sputum, or facility infrastructure.

AI-enabled mobile phone applications can support continuous and automated cough monitoring while on tuberculosis (TB) treatment. This has potential to serve as an early warning system to identify individuals at risk of unfavorable treatment outcomes.Methods: We enrolled adults ≥18 years in Kampala, Uganda who were initiated on treatment for drug-sensitive pulmonary TB based on a microbiological (sputum Xpert MTB/RIF) or clinical diagnosis. We collected clinical and sociodemographic data, and participants wore smartphones for 14 days with the Hyfe application for automated cough detection. We calculated the daily median coughs per hour and applied an anomaly detection model to identify individuals with an extreme cough frequency. TB registry data were used to identify unfavorable outcomes (Death or loss to follow up). We developed a 10-fold cross validated LASSO regression model to predict unfavorable outcomes and calculated the accuracy.Results: Among 152 participants, median age was 32 years (IQR 40-26), 66 (43.42%) were living with HIV, 94 (61.84%) had bacteriologically confirmed TB, and 6 (3.9%) had an unfavorable outcome (3 deaths and 3 lost to follow up). Participants with unfavorable outcomes had a higher median cough per hour on Day 1 than those who were cured or completed treatment, but levels were similar by Day 14 (Figure 1). A Day 1 cough frequency anomaly classified an unfavorable outcome with 83% sensitivity (95% CI 45-100) and 78% specificity (95% CI 25-85). When combined with presence of hemoptysis, microbiological TB status, and history of HIV, diabetes or allergies, sensitivity remained 83% and specificity improved to 94%.Conclusions: Continuous automated cough monitoring in the first 24 hours identified individuals with unfavorable outcomes with high accuracy when combined with routine baseline clinical data and should be further evaluated as a TB treatment monitoring tool in larger validation studies.