AI stethoscope shows promise for tuberculosis screening

Tuberculosis remains a major global killer: in 2023 there were an estimated 10.8 million cases and 1.3 million deaths, and roughly 2.6 million cases were missed by existing screening systems. Current approaches such as symptom-based screening and chest X-ray (CXR) with computer-aided detection (CAD-CXR) can be costly, hard to access, or not sensitive enough to catch all infections. In response, researchers led by corresponding author Max Rath explored whether sounds from the chest, captured with a digital stethoscope and interpreted by artificial intelligence, could be an alternative screening method. Between June 2021 and November 2022, trained nurses in South Africa’s Western Cape province used AI Diagnostics' prototype digital stethoscope to record lung sounds. The study collected 49,770 anonymized chest auscultation recordings from 1,659 participants recruited at 34 primary care clinics; participants were people suspected to have TB who reported a recent sputum TB Xpert Ultra test. The goal was to see if AI could learn to recognise lung sound patterns that indicate pulmonary TB, offering a portable, lower-cost option for screening in settings where X-rays and laboratory tests are less available.

After removing and stratifying data as needed, the research team prepared a final dataset of 1,169 participants. That dataset was split into an 80% training set and a 20% hold-out test set. The team fine-tuned a pre-trained transformer-based architecture using K-fold cross-validation and combined predictions into an ensemble model. The ensemble's performance was evaluated on the hold-out set with sputum TB Xpert Ultra used as the reference standard. On that independent test set the model achieved a mean Area under the Receiver Operating Curve (AUC-ROC) of 0.79 (95% CI: 0.73-0.85). At a high sensitivity setting the model reached 89.9% sensitivity (95% CI: 82.4%-94.4%) while producing 50.4% specificity (95% CI: 42.0%–58.7%) for predicting pulmonary TB from lung sounds. The study was commercially funded by AI Diagnostics Pty (Ltd).

These results suggest AI-based digital chest auscultation could be a useful adjunct or alternative to existing TB screening methods, particularly where cost, equipment, or access limit the use of CXR and laboratory tests. A sensitivity near 90% indicates the approach may catch many true cases, including people who do not report symptoms and would be missed by symptom checks alone, though the modest specificity means many people without TB would still be flagged for further testing. The technology's portability and lower cost could allow screening in remote clinics and community settings and help bridge gaps in places with high TB burden. The authors note that these findings are early and that future independent studies in diverse, unselected populations with high TB prevalence are needed to confirm how well the model generalises. If validated independently, this approach might inform clinical practice, national screening programmes, and global guidance, but the study's commercial funding underlines the importance of external replication and transparent assessment before wide deployment.