Abstract
BACKGROUND: Chest radiography (CXR) is the most widely used imaging tool in pediatric tuberculosis (TB) diagnostic pathways, and remains central in current WHO algorithms. However, its standalone diagnostic accuracy has not been well established in standardized multicenter evaluations. This study aimed to determine the diagnostic performance and interobserver agreement of CXR for pediatric TB across two epidemiologically distinct settings, and to assess the added value of clinical information and lateral projections. METHODS: We evaluated the diagnostic performance of CXR in two pediatric cohorts from distinct TB-burden settings. The high-burden cohort (Mozambique) included 218 children under 3 years (10 confirmed TB, 95 unconfirmed TB, 113 unlikely TB). The low-burden cohort (Spain) included 674 children under 18 years (145 confirmed TB, 237 unconfirmed TB, 95 with TB infection, 101 with community-acquired pneumonia, and 96 healthy controls). Four independent expert readers (three pediatric radiologists and one pediatric infectious disease specialist), each with over 15 years of experience, interpreted CXRs using a standardized digital platform, blinded to clinical data. In a subset of 75 Spanish cases, re-readings incorporated limited clinical information. RESULTS: Sensitivity for confirmed TB was low in both settings (31.0% in Mozambique, 46.1% in Spain), while specificity was high (94.7% and 96.5%, respectively). In a subset of 75 Spanish cases, adding limited clinical data increased sensitivity from 39.3% to 50.0% (p = 0.02) and specificity from 88.1% to 97.4% (p < 0.001). Among children with lateral views, sensitivity rose from 39.1% to 53.6% (p = 0.01), without significant change in specificity. Interobserver agreement for TB-related findings was only fair (ICC 0.29-0.31). CONCLUSIONS: This multicenter analysis confirms the limited sensitivity but high specificity of CXR for pediatric TB, even when interpreted by expert readers. These findings highlight that CXR alone cannot reliably confirm or exclude disease and should be integrated with clinical and microbiological data. Future diagnostic pathways, including artificial intelligence-assisted CXR interpretation, will likely need multimodal approaches to overcome the intrinsic limitations of imaging alone.