Investigators analyzed a pediatric symptomatic cohort of 452 children aged 15 years or younger in Nairobi. The group underwent a standard diagnostic panel: chest x-ray (CXR), respiratory sampling tested by GeneXpert MTB/RIF Ultra (Xpert Ultra), smear microscopy and culture, plus immune-based assays — tuberculin skin test (TST) and interferon gamma release assay (IGRA). The authors reported culture as the conventional reference, but noted its limited sensitivity in children; they cited expected culture sensitivity ranges rather than producing new culture-only gold-standard results. The manuscript is a preprint and has not undergone peer review.
Three Bayesian latent class models were fitted to these data. Model 1 assumed conditional independence across all diagnostic tests. Model 2 introduced random effects to capture dependence within biologically related tests (smear, Xpert Ultra and culture) and within immunologic tests (IGRA and TST). Model 3 extended Model 2 by incorporating clinical covariates and NIH-defined TB signs and symptoms to adjust prevalence and test performance estimates. The study team favored Model 3 because it modeled known sources of dependence among assays and adjusted for clinical predictors, thereby producing prevalence and accuracy estimates that account for both test correlation and patient-level factors.
The models yielded divergent prevalence estimates. Model 1 returned a posterior prevalence estimate of approximately 20.2% (15.9–24.7). Model 2 produced a higher estimate near 24.5% (19.6–30.0). When clinical covariates were added in Model 3, estimated prevalence fell to about 14.8% (11.0–19.3). All three model-derived prevalence estimates exceeded the 9.9% prevalence calculated from microbiological confirmation alone. The authors therefore suggested that reliance on culture/Xpert/smear alone underestimates the burden of pediatric tuberculosis in this symptomatic clinic population.
Under the authors’ preferred Model 3, the molecular and microbiologic assays showed moderate sensitivity with high specificity. Culture sensitivity was estimated at 56.9%; GeneXpert MTB/RIF Ultra sensitivity at 63.2%. Smear microscopy sensitivity was substantially lower at 29.8%. Specificity for culture and Xpert Ultra remained high (authors described excellent specificity; exact numeric ranges are reported in the source). Radiography achieved a moderate sensitivity (79.7%) but poor specificity (40.2%), which limits its capacity to rule in disease despite utility as a screening tool. Immune-based diagnostics had intermediate performance: IGRA sensitivity 63.7% with specificity 81.5%; TST sensitivity 50.1% with specificity 65.8%.
The report emphasized that sensitivity and specificity estimates were broadly consistent across the three models, but that absolute prevalence and some accuracy estimates shifted when dependence and covariates were modeled.
Model 3 included clinical covariates and NIH TB signs and symptoms. The authors observed a positive association between older age and modeled TB probability: the log-odds ratio for age was 1.09 (95% credible interval 0.14–2.18) in the adjusted model. The source did not provide exhaustive detail on all covariates’ coefficients in the abstract; full parameter estimates appear in the primary report.
The study team performed sensitivity analyses to assess influence of prior distributions. They reported that prevalence and accuracy estimates were robust to changes in priors. Model 3 was selected as preferred primarily because it explicitly accounted for dependence between tests and adjusted for clinical covariates, which altered prevalence and certain sensitivity estimates relative to simpler models.
The authors confirmed ethical approvals from the University of Nairobi, Kenyatta National Hospital and the Kenya Medical Research Institute Scientific and Ethics Review Unit. They noted that underlying data include sensitive pediatric patient information (including HIV status) and are subject to institutional restrictions; de-identified data may be available on reasonable request to the named contact. The manuscript is a medRxiv preprint and has not been peer reviewed; the authors present their findings as preliminary. The abstract reports limited methodological detail (for example, the specific covariates included and full posterior intervals for all parameters are not enumerated in the abstract), so readers should consult the full preprint for complete model specifications and comprehensive numerical results.