TB scoring tool caught 97% of cases in children with HIV, four-country trial finds

Tuberculosis is difficult to confirm in a young child, and it is deadliest in the children hardest to test. A diagnostic cohort study published in The Lancet Global Health on September 14 set out to see whether a simple scoring tool could close that gap for children living with HIV — and reports that it largely did [s1].

The World Health Organization's tuberculosis fact sheet frames the stakes. TB is the leading cause of death among people with HIV, people living with HIV are 12 times more likely to fall ill with TB disease than people without it, and in 2024 about 150,000 people died of HIV-associated TB [s2].

What was tested

The tool is the PAANTHER treatment decision algorithm, developed by the Pediatric Asian African Network for Tuberculosis and HIV Research. The TB-Speed HIV study set out to externally validate it in a prospective cohort of children in Africa [s1].

The study ran in seven tertiary hospitals — two in Côte d'Ivoire, two in Mozambique, one in Uganda, and two in Zambia — enrolling children living with HIV aged 1 month to 14 years with presumptive tuberculosis [s1]. At enrolment each child was assessed for suggestive symptoms and underwent Xpert MTB/RIF Ultra testing on respiratory and stool samples, chest radiography, and abdominal ultrasonography [s1]. Those inputs produced a PAANTHER score: below 100 was negative, while 100 or higher was positive and prompted the start of treatment [s1].

At the end of the study an endpoint review committee retrospectively classified each child as having confirmed, unconfirmed, or unlikely tuberculosis. That classification formed the composite reference standard against which the algorithm was judged [s1].

The numbers

Between October 2, 2019, and December 31, 2021, 1,706 children were prescreened, 713 were screened for tuberculosis symptoms, 423 were eligible, and 277 were enrolled [s1]. Of those 277, 272 (98%) had a complete algorithm evaluation: 215 (79%) scored 100 or higher, including 24 (9%) with a positive Xpert MTB/RIF Ultra test [s1].

Treatment was initiated in 185 (86%) of the 215 children with a positive score and in 12 (20%) of the 60 children with a negative score, at a median of one day after inclusion (IQR 0–4) [s1].

The disease was common in this group. After the committee's assessment, the proportion of children with tuberculosis in the classifiable population of 273 was 56.8% (95% CI 50.9–62.5), with 131 classified as having unconfirmed and 24 as having confirmed tuberculosis [s1].

The headline safety measure was how many cases the tool cleared in error. Four (7%) of the 60 children with a negative score were subsequently classified by the committee as having tuberculosis — the missed cases [s1]. Set against the composite reference standard, the algorithm had a negative predictive value of 93.3% (95% CI 84.1–97.4) and a sensitivity of 97.4% (95% CI 93.6–99.0) [s1]. The protocol had defined a negative predictive value of 75% as the threshold for validation, so the result cleared it comfortably [s1].

The trade-off sits on the other side of the ledger. Positive predictive value was 70.9% (95% CI 64.5–76.6) and specificity just 47.5% (95% CI 38.7–56.4) [s1]. A tool this sensitive inevitably flags children who do not have the disease; roughly three in ten positive scores were false positives, and fewer than half of the children without tuberculosis were correctly ruled in as negative.

Why the balance was chosen

That is a deliberate design choice, not a defect. In a child living with HIV, a missed tuberculosis diagnosis can be fatal within weeks, while a false positive costs a course of treatment. An algorithm tuned to miss almost nothing will, by construction, over-treat. The authors conclude that the high sensitivity of the tool and its potential to enable rapid treatment initiation could contribute to reducing mortality in children living with HIV [s1].

The speed matters as much as the accuracy. A median of one day from assessment to treatment is the kind of interval that a microbiological work-up, dependent on culture, rarely achieves in these settings — which is the whole point of a scoring tool built from symptoms, a rapid molecular test, and imaging rather than from confirmation.

What it does not settle

The study validates the algorithm against a reference standard that is itself a committee judgment, not a bacteriological gold standard, because no reliable gold standard exists in young children. The composite includes a large "unconfirmed" category — 131 of the classified children — which is a clinical designation, not a laboratory one [s1]. The trial was conducted in tertiary hospitals with access to Xpert Ultra, chest radiography, and ultrasonography [s1], and whether the same performance holds where those inputs are thinner is a separate question the paper does not answer.

The wider treatment gap remains large. Globally in 2024, only 61% of the estimated number of people living with HIV who developed tuberculosis received antiretroviral therapy, by WHO's account [s2]. A validated tool to find paediatric cases faster is one piece of a system that still misses a great many of them.

The study was registered as NCT04121026 and funded by Unitaid [s1].

This article is informational and is not medical advice.