Diagnostic accuracy of a commercial AI digital stethoscope for diagnosis of TB.

Background: Improved TB screening requires non-invasive, low-cost, and rapid diagnostics. Digital stethoscopes utilising machine-learning approaches to analyse respiratory sounds have potential.

Methods: We assessed accuracy of a commercial digital stethoscope for TB diagnosis among TB symptomatic participants. The microbiological reference standard (MRS) was sputum TB-positive on either liquid culture, solid culture, or Xpert MTB/RIF Ultra. Adults were enrolled from South Africa, Uganda, Vietnam, and Peru, with pre-defined sampling of 60 MRS-positive and 180 MRS-negative participants over two stages. Respiratory sounds from six auscultation positions on the participant's torso were analysed. The manufacturer (blinded to MRS status) provided participant scores and a test-positivity cut-off.

Results: Among 240 participants, 135 (56%) were female, 62 (26%) living with HIV, 35 (15%) current smokers, and 31 (13%) previously treated for TB. Estimates of sensitivity and specificity, adjusted for country-stratified sampling, were 77% (95% confidence interval [CI]: 65-85) and 50% (95% CI: 43-57), respectively. Sensitivity was lower among people living with HIV and those with sputum smear-negative TB and varied by country. Testing took 5 min per participant (median, interquartile range 4-6).

Conclusion: These early data suggest that further refinement of this test is warranted. The device is simple to use, is inexpensive, and can be used offline.