Abstract
Objectives
Tuberculosis (TB) remains a major health threat in prisons, where overcrowding and limited access to care facilitate transmission. We evaluated the diagnostic accuracy and feasibility of the AeoNose for TB screening in Paraguayan prisons.
Methods
In this cross-sectional diagnostic study, people deprived of liberty and prison staff underwent symptom screening, chest X-ray with CAD4TB (v6), and AeoNose breath sampling. Individuals with presumptive TB underwent GeneXpert and mycobacterial culture. Breath signals were analyzed using a convolutional neural network. Two subsets were evaluated: all participants with valid samples and the same population excluding prior TB. Five-fold cross-validation was applied.
Results
Of 3752 individuals screened, 2993 had valid samples. TB prevalence was 6.7%. In subset 1 (188 TB cases) and subset 2 (142 TB cases), mean AUC was 0.634 (SD 0.040) and 0.624 (0.054). Sensitivity and specificity were 0.670 (95% CI: 0.592-0.737) and 0.537 (0.518-0.555), and 0.711 (0.630-0.793) and 0.505 (0.491-0.519).
Conclusion
Under real-world prison conditions, AeoNose did not meet WHO criteria for a TB triage test. However, large-scale breath-based screening using portable devices was feasible, supporting further development of non-invasive screening approaches.