AI chest X-ray tuberculosis screening nears the WHO triage bar, but consistency remains uneven
Artificial intelligence for pulmonary tuberculosis detection on chest X-ray is moving closer to the performance level needed for triage in crowded screening programs, yet the margin is still thin enough to matter at the bedside and in public health workflows. In this systematic review and diagnostic accuracy meta-analysis, investigators pooled clinical cohort data to judge whether AI-assisted chest radiography can reliably separate people who need further testing from those who do not. The answer is encouraging on sensitivity and much less settled on specificity.
Nine evaluations from three independent cohorts were suitable for pooled analysis. Across those data, pooled sensitivity reached 0.896, with a 95% confidence interval of 0.874 to 0.915 and a prediction interval of 0.798 to 0.950. Pooled specificity was 0.689, with a 95% confidence interval of 0.642 to 0.732 and a prediction interval of 0.499 to 0.831. Both confidence intervals sit near the World Health Organization triage thresholds of 90% sensitivity and 70% specificity, but the prediction interval for specificity stretches from clearly subthreshold to comfortably above target, a spread that signals real-world instability across cohorts and products.
One cohort can look reassuring; three cohorts can still disagree.
That variability is the central clinical problem. A triage tool that misses too many cases weakens tuberculosis control, but a tool with modest specificity can flood referral pathways with false positives, increasing sputum testing, imaging, clinic visits, and patient anxiety. For tuberculosis programs already strained by radiologist shortages and high case burden, a false-positive-heavy algorithm does not merely inconvenience staff. It shifts workload downstream and can crowd out patients who need microbiologic confirmation most urgently.
Clinical cohorts, not neat test sets, exposed the performance gap
The review deliberately focused on evaluations drawn from clinical cohorts with complete 2 × 2 data and a microbiological or composite reference standard applied to all participants. That choice makes the analysis more relevant to actual care settings than studies built around curated image collections. Five databases were searched on 15 March 2026 for studies published between January 2020 and March 2026, and QUADAS-2 was used to appraise risk of bias. Pooled estimates were generated with a bivariate random-effects model, and 95% prediction intervals were included to show how much performance could vary in future settings.
Only 21 studies, representing 26 evaluations of AI models, met inclusion criteria. The evidence base narrowed further after accounting for overlapping participants, leaving one evaluation per cohort in a post hoc analysis. A screening study with partial verification was added in a secondary analysis, and pooled specificity increased to 0.735. Even so, the evidence remained anchored in only three independent cohorts. That is a narrow foundation for broad implementation decisions, especially when tuberculosis screening often occurs in settings where prevalence, HIV co-infection, age structure, image quality, and referral behavior differ sharply from one site to the next.
The performance spread likely reflects more than model design alone. Threshold selection changes the balance between sensitivity and specificity, and local disease prevalence can influence how a given operating point behaves in practice. Image acquisition quality, portable versus stationary radiography, reader workflow, and the mix of smear-positive and smear-negative disease can all shape apparent accuracy. In tuberculosis, radiographic patterns can be subtle, especially when cavitation is absent or disease is early, disseminated, or superimposed on prior lung injury. AI systems trained or calibrated in one environment may therefore travel poorly unless their decision thresholds are adapted carefully.
Bias, verification, and external validation limit confidence
The review does not present a clean endorsement of current AI-CXR products. Eighteen of 24 assessed evaluations were at high risk of bias, and only one study was prospective, fully verified, externally validated, and met the WHO thresholds. Certainty of evidence was rated low for sensitivity and very low for specificity. Those ratings are hard to ignore because diagnostic accuracy claims often look stronger in development datasets than they do in routine practice, where incomplete verification and selective inclusion can inflate apparent performance.
Partial verification is especially troublesome in tuberculosis screening. If only some individuals receive the reference standard, specificity can look better or worse depending on who was worked up further. External validation matters for the same reason: an algorithm that performs well in one hospital or country may not preserve that behavior elsewhere. The review’s own framework captures these concerns by emphasizing clinical cohorts and complete reference standards, yet the resulting evidence still depends on a small number of cohorts and a heterogeneous set of AI products.
A subsequent PubMed search found additional eligible clinical evaluations that were not screened, so the review is incomplete for those studies. That introduces a practical limitation in the evidence map itself, even before considering the usual caveats about a preprint that has not undergone peer review. The missing studies could alter the pooled picture, especially if they come from underrepresented regions or include newer algorithms with different calibration characteristics.
Local calibration will determine whether AI helps or clutters the pathway
For tuberculosis programs, the key question is not whether AI can read a chest X-ray at all. The real issue is whether it can do so at a threshold that preserves scarce confirmatory testing for the people most likely to benefit. The meta-analysis suggests the answer may depend on where and how the system is used. Sensitivity is close to the WHO triage target, but specificity is only close to target and not clearly above it across settings.
Local threshold calibration will therefore be essential before broad deployment. A program with a high tuberculosis burden may accept lower specificity if the aim is to minimize missed cases, while a setting with constrained diagnostic capacity may need a more conservative operating point to avoid overwhelming follow-up services. Because the prediction intervals are wide, each program will need to validate the algorithm against its own patient mix, image quality, and confirmation pathway rather than assuming pooled performance will transfer intact.
The strongest clinical message is restrained, not triumphant: AI-assisted chest X-ray is approaching the level where it may support tuberculosis triage, but the evidence remains too narrow, too heterogeneous, and too dependent on setting to justify blanket deployment. Prospective, externally validated evaluations in real clinical cohorts are still needed, and the current preprint provides no shortcut around that requirement.