Validation of a Paediatric-Optimized Computer-Aided Detection System for Tuberculosis Using Bayesian Latent Class Analysis

Background Microbiological confirmation of paediatric pulmonary tuberculosis is frequently unattainable, rendering chest radiography a critical yet underutilised diagnostic tool. Methods We conducted a retrospective diagnostic accuracy study of the qXR version 4.2.1 (Qure.ai), a paediatric optimized computer-aided detection (CAD) algorithm, for pulmonary tuberculosis. Diagnostic performance was assessed against microbiological (MRS) and clinical reference standards (ClRS). Bayesian latent class analysis (LCA) was applied to address the imperfection of both reference standards in children. Performance was quantified using area under the receiver operating characteristic curve (AUROC) and estimates of sensitivity and specificity. Results We included digital chest radiographs of 932 Gambian children (< 15 years) comprising 80 (9%) children with confirmed tuberculosis, 163 (17%) with unconfirmed tuberculosis, and 689 (74%) classified as unlikely tuberculosis. Against MRS, qXR demonstrated AUROC, sensitivity and specificity of 0.68 (95% CI, 0.61 to 0.75), 54% (95% CI, 43 to 64%), and 82% (95% CI, 79 to 84%), respectively. Against ClRS, the AUROC, sensitivity and specificity were 0.73 (95% CI, 0.69 to 0.77), 41% (95% CI, 34 to 49%), and 87% (95% CI, 84 to 89%), respectively. Bayesian LCA, assuming conditional independence, estimated sensitivity of 79% (95% CrI, 65 to 89%) and specificity of 82% (95% CrI, 79 to 84%). Assuming conditional dependence between qXR and expert radiologist, and between culture and Xpert, estimated sensitivity increased to 89% (95% CrI, 71 to 98%), with specificity remaining at 82% (95% CrI, 79 to 84%). Conclusions Paediatric optimized qXR algorithm provides a valuable complementary tool for diagnosis of paediatric pulmonary tuberculosis. Conventional reference standards likely underestimate the true diagnostic performance of CAD systems in children.