A systematic review and meta-analysis shows that artificial intelligence (AI) can detect tuberculosis (TB) from cough recordings with high accuracy. The pooled sensitivity was 91% and the specificity was 89%. The authors see this as a potentially low-cost and easily accessible screening tool for resource-limited settings. The study was published in the journal Frontiers in Artificial Intelligence.
Background
Tuberculosis is the most common fatal infectious disease worldwide. Especially in low- and middle-income countries, there is often a lack of rapid, inexpensive, and readily available diagnostic methods. The typical TB cough differs acoustically from coughs in other diseases. AI-supported analysis of cough recordings could therefore represent a non-invasive, scalable screening method – especially with the help of smartphones.
Method and Results
The researchers analyzed 14 studies (mainly from Asia and Africa) published between 2009 and 2024. In seven studies with sufficient data for a meta-analysis, the AI achieved a pooled sensitivity of 91% (95% CI: 88–94%) and a specificity of 89% (95% CI: 85–92%). The area under the ROC curve (AUC) was 0.9539 – a very good value.
Deep learning models performed slightly better (sensitivity 92%, specificity 91%) than classical machine learning approaches. Most studies used features such as Mel-Frequency Cepstral Coefficients (MFCCs) or spectrograms and trained models like ResNet, VGG, or LSTM architectures.
Assessment of Study Quality
Despite the good statistical results, the methodological quality of the studies was predominantly limited. Many studies showed a high risk of bias (especially in patient selection). Most studies were purely analytical validations on existing datasets and not tested in real-world clinical workflows. Large, prospective multicenter studies with external validation are still lacking.
Furthermore, the heterogeneity between the studies was high, and there was evidence of publication bias (smaller studies with better results were more frequently represented).
Significance and Outlook
The results show that AI-based cough analysis could be fundamentally suitable for early screening of tuberculosis in resource-limited regions – quickly, cheaply, and without laboratory equipment. However, the authors caution against adopting the technology into screening programs already. Before widespread use is sensible, prospective clinical validation studies under real-world conditions must be conducted.
FAQ
How does AI detect TB from coughs?
The AI analyzes acoustic features of the cough (e.g., frequency patterns, temporal progressions) that can be characteristically altered in tuberculosis.
How good is the method?
In the meta-analysis, it achieved a sensitivity of 91% and specificity of 89%. This is very good for a screening method – but under controlled study conditions.
Can this already be done with a smartphone?
Technically, yes – many studies used smartphone recordings. However, robust, clinically validated systems are still lacking for real-world use.
Why is the method particularly interesting for poor countries?
It would be very inexpensive, non-invasive, and could also be used in remote regions without laboratory infrastructure.
When could the technology arrive in practice?
The authors currently do not see it as ready for routine use. Prospective multicenter studies with external validation are urgently needed.
