# Diagnostic accuracy of the WHO tuberculosis treatment decision algorithms for children with presumptive tuberculosis: An individual participant data meta-analysis

## TL;DR

This study evaluates WHO's tuberculosis treatment decision algorithms for children using real-world data, finding high sensitivity but low specificity, which could lead to over-treatment.

## Contribution

The study is the first to validate WHO's treatment decision algorithms using individual participant data from multiple cohorts, including vulnerable populations.

## Key findings

- The algorithms showed high sensitivity (84.3% for TDA A and 90.6% for TDA B) but low specificity (50.6% for TDA A and 30.8% for TDA B).
- Sensitivity was lower in high-risk groups like children under 2 years, those with HIV, or malnutrition.
- The study confirms the algorithms' performance in real-world settings but highlights the need for improved specificity to avoid over-treatment.

## Abstract

In 2023, almost 200,000 children under 15 years died from tuberculosis, most without appropriate treatment. Treatment decision algorithms (TDAs), developed to facilitate rapid anti-tuberculosis treatment initiation in children, were recommended by the World Health Organization (WHO) in 2022, conditional on validation in different cohorts and settings. We performed a retrospective external evaluation of WHO TDAs using an individual participant dataset (IPD).

The IPD comprised four paediatric cohorts, restricted to children with presumptive pulmonary TB < 10 years, and including children in high-risk groups (children living with HIV “CLHIV”, children with severe acute malnutrition “SAM”, and children <2 years). All children in the IPD were retrospectively evaluated using both TDA A (an algorithm including chest X-ray) and TDA B (without chest X-ray), excluding the triage step. The diagnostic accuracy against a composite reference standard (confirmed and unconfirmed tuberculosis versus unlikely tuberculosis) was determined and reported as sensitivities and specificities. Of 1,886 children included (RaPaed-TB: n = 740, Umoya: n = 474, TB-Speed HIV: n = 204, TB-Speed Decentralisation: n = 468), the median age was 2.9 years (interquartile range [IQR]:1.3,5.5), 741 (39.3%) were <2 years, 382 (20.3%) were CLHIV, and 284 (15.1%) had SAM. 281 (14.9%) had confirmed tuberculosis, 672 (35.6%) were classified as unconfirmed tuberculosis (clinically diagnosed, microbiological investigations negative), and 933 (49.5%) as unlikely tuberculosis. For TDAs A and B, algorithm sensitivity was 84.3% (95% CI: 74.8, 90.6) and 90.6% (95% CI: 83.8, 94.7), respectively, with a specificity of 50.6% (95% CI: 30.4, 70.7) and 30.8% (95% CI: 21.5, 42.0), respectively. For TDA A, estimated sensitivity in children in high-risk groups was lower than those with low-risk (83.0%, 95% CI: 79.4%, 86.1%; versus 88.0%, 95% CI: 84.8%, 90.6%), while having a gain in specificity (50.0%, 95% CI: 44.9%, 55.1%; versus 36.6%, 95% CI: 32.7%, 40.7%). Trends were similar for TDA B. As for limitations, most diagnostic tuberculosis studies in children, including two of those included in the IPD, are performed at secondary or tertiary hospitals with higher levels of healthcare and thus the target population might differ somewhat from the IPD, potentially limiting the generalisability of our results.

This retrospective external evaluation of WHO TDAs in a large IPD shows high sensitivity but sub-optimal specificity for both TDAs, in line with the meta-analyses that generated the algorithms. Prospective studies that evaluate the entire TDA, including triage step are needed. Additionally, the integration of novel diagnostic tools within the TDAs should aim to enhance the accuracy, especially the specificity.

Tuberculosis in children remains one of the top 10 causes of death in those younger than 5 years, mainly due to missed or delayed diagnosis. This is especially challenging in primary healthcare settings, where available tests are difficult to perform, require substantial infrastructure, and lack sufficient accuracy.

In 2022, WHO recommended treatment decision algorithms for TB, which are simple flow-charts designed to guide healthcare workers step by step through a standardised diagnostic process that relies primarily on clinical information. These algorithms aim to support and standardise treatment decision, but evidence on their performance remains limited.

In our study, we used data from several large previously conducted studies on children that underwent testing for tuberculosis, to evaluate how well these treatment decision algorithms perform to identify children with tuberculosis.

We found that the performance in this independent dataset of children was comparable to that reported in the original discovery study. While the algorithms identified a large number of children with tuberculosis (high sensitivity), it also recommended to start a considerable number of children without tuberculosis on treatment (sub-optimal specificity).

The accuracy was also similar in those children of vulnerable populations, including young children and those affected by HIV or malnutrition.

To our knowledge, this is the first study to use previously collected data from several studies with individual participant datasets to assess the accuracy of WHO treatment decision algorithms. We validate the estimated performance using real-world data, importantly confirming its accuracy in vulnerable populations. However, low specificity might lead to substantial overtreatment, underscoring the urgent need for novel diagnostic tools with higher specificity.

Our findings underscore the potential usefulness of diagnostic approaches such as treatment decision algorithms to identify more children eligible for tuberculosis treatment. By using a tool that can potentially be implemented at low levels of healthcare, this approach might help to avert many deaths due to childhood tuberculosis.

Limitations include the heterogeneity of studies, partially conducted at higher levels of care, which may limit applicability to broader populations. In addition, due to the retrospective nature of the study, the initial triage/ screening step couldn’t be assessed.

Laura Olbrich and colleagues use real-world individual participant data to assess the diagnostic accuracy of WHO tuberculosis treatment decision algorithms that aim to accelerate anti-tuberculosis treatment initiation in children.