Beyond accuracy: evaluating the operational feasibility and diagnostic yield of CAD4TB vs. Timika score for scalable TB screening in low-resource settings.

Artificial intelligence has shown promise in enhancing tuberculosis care, but its use in resource-limited settings like Indonesia remains underexplored. This cross-sectional retrospective single-centre study evaluates the diagnostic performance of CAD4TB in screening Indonesian patients suspected of tuberculosis using chest x-ray (CXR) images, comparing its efficacy to the Timika score assessed by experts. We analyzed CXR images from 3,254 patients (2018-2020), including 600 with microbiological confirmation, of whom 46 had smear-positive pulmonary tuberculosis (PTB). CAD4TB demonstrated an area under the curve (AUC) of 0.778 (95% CI 0.712-0.844) when compared to acid-fast bacilli (AFB) results without a time interval. With a ≤7-day interval between CXR and AFB data, CAD4TB showed an AUC of 0.767 (95% CI 0.668-0.866), comparable to the Timika score of 0.726 (95% CI 0.632-0.820). Additionally, CAD4TB exhibited superior specificity (71.43% vs. 57.64%, p < 0.001) while maintaining a fixed sensitivity of 73.91%. These findings suggest that CAD4TB outperforms the Timika score and holds promise as a rapid tuberculosis screening tool in resource-limited settings like Indonesia.