Predicting radiological severity of pulmonary tuberculosis in children: an assessment of the WHO-criteria and novel prediction scores on an individual participant dataset

Technical Summary: Predicting Radiological Severity of Pulmonary Tuberculosis in Children

Problem Statement
The World Health Organization (WHO) recommends a shortened 4-month antituberculosis regimen for children with non-severe pulmonary tuberculosis (TB), a strategy validated by the SHINE trial. However, the eligibility criteria for this regimen rely on the exclusion of radiologically severe disease, typically confirmed via chest X-ray (CXR). In many resource-limited settings, CXR access is unavailable. The WHO has proposed alternative eligibility criteria ("WHO-criteria") for settings without CXR, based on clinical symptoms and bacteriological testing status. However, these criteria are based on expert consensus rather than empirical evidence, raising concerns about their ability to accurately identify children with severe disease who require the standard 6-month regimen. Misclassification risks undertreatment (leading to poor outcomes) or overtreatment (increasing drug toxicity and costs). There is a critical need to assess the diagnostic accuracy of these existing criteria and develop data-driven clinical prediction scores to support treatment decisions in the absence of radiology.

Methodology
This study utilized an Individual Participant Dataset (IPD) from the Decide-TB project, aggregating data from four diagnostic studies (RaPaed-TB, Umoya, TB-Speed HIV, and TB-Speed Decentralisation) across eight countries. The final analysis included 2,383 children, of whom 633 met the inclusion criteria for the 4-month regimen assessment (ages 3 months to 10 years, excluding those with severe clinical presentations like severe acute malnutrition or advanced HIV).

• Reference Standard: Radiological severity was determined by a single independent pediatric TB expert reviewing baseline CXRs. "Radiologically non-severe" was defined as normal CXRs or specific abnormalities (e.g., unilateral lymphadenopathy, single-lobe consolidation without cavities). "Radiologically severe" included multilobe disease, cavities, miliary patterns, airway obstruction, or pleural complications.
• Evaluation of WHO-Criteria: The diagnostic accuracy of the WHO-criteria (both with and without bacteriological testing) was assessed against the expert CXR reference standard. Sensitivity and specificity were calculated using bootstrapping (1,000 iterations).
• Prediction Score Development: Two multivariable logistic regression models with random intercepts (to account for study heterogeneity) were developed:
  • Model 1 (Score 1): Included 11 clinical predictors plus bacteriological testing results (Xpert MTB/RIF or Ultra semi-quantitative results or smear microscopy).
  • Model 2 (Score 2): Included the same 11 clinical predictors but excluded bacteriological testing.
  • Predictors included age, sex, weight/BMI-for-age Z-score, cough duration, fever duration, TB exposure history, HIV status, night sweats, lymphadenopathy, tachycardia, and tachypnea.
  • Models were internally validated using bootstrap resampling. Coefficients were converted into integer scoring systems (Score 1 and Score 2) with a target sensitivity of ≥70% to minimize missed severe cases.
• Analysis: Diagnostic performance (sensitivity, specificity, AUROC) was evaluated for the WHO-criteria and the new scores across various subgroups (age, HIV status, nutritional status, exposure history).

Key Results

• Population Characteristics: Of the 633 children analyzed, 26.6% (168/633) met the initial eligibility criteria for the 4-month regimen. Among these, 18.3% (116/633) were found to have radiologically severe disease upon expert review.
• Performance of WHO-Criteria:
  • With Bacteriological Testing: Sensitivity was 30.1% (95% CI: 20.3–40.2%) and specificity was 83.4% (95% CI: 80.2–86.4%).
  • Without Bacteriological Testing: Sensitivity dropped to 21.7% (95% CI: 10.4–34.5%) with a specificity of 81.9% (95% CI: 78.8–84.9%).
  • The criteria showed consistently low sensitivity across subgroups, particularly in children aged 2–5 years and those with a history of TB exposure.
• Performance of Developed Scores:
  • Score 1 (With Bacteriology): Achieved a sensitivity of 41.1% (95% CI: 32.4–49.5%) and specificity of 77.3% (95% CI: 73.6–80.8%). The optimism-corrected AUROC was 0.708.
  • Score 2 (Without Bacteriology): Achieved a sensitivity of 30.9% (95% CI: 22.6–40.4%) and specificity of 83.0% (95% CI: 79.5–86.3%). The optimism-corrected AUROC was 0.676.
• Risk of Undertreatment: Using the WHO-criteria with bacteriology, 78.4% (91/116) of children with severe disease were incorrectly classified as eligible for the 4-month regimen (risk of undertreatment). This improved to 58.6% (68/116) when using Score 1. Without bacteriology, the risk of undertreatment was 90.5% with WHO-criteria and 68.9% with Score 2.

Key Contributions

1. First Empirical Assessment: This is the first study to empirically evaluate the diagnostic accuracy of the WHO eligibility criteria for the 4-month regimen in settings without CXR, revealing significant limitations in their ability to detect severe disease.
2. Development of Prediction Scores: The authors developed and validated two novel clinical prediction scores (Score 1 and Score 2) designed to complement WHO Treatment Decision Algorithms (TDAs) in resource-limited settings.
3. Quantification of Risk: The study quantifies the substantial risk of undertreatment associated with current guidelines, demonstrating that while the new scores improve sensitivity over the WHO-criteria, they remain insufficient for reliable standalone use.

Significance and Claims
The paper concludes that while the developed scores demonstrate better sensitivity than the current WHO-criteria, their overall performance remains suboptimal for reliably identifying children with radiologically severe tuberculosis based on clinical features alone. The authors state that implementing the 4-month regimen in the absence of adequate CXR infrastructure remains challenging. They emphasize that the current data-driven approach cannot fully replace the need for radiological assessment. Consequently, the paper advocates for improved access to affordable and feasible CXR in resource-limited settings and suggests that computer-aided detection tools or further research into alternative severity classification methods are necessary to safely implement shortened regimens without radiology. The authors caution that their findings highlight a need for careful clinical follow-up and do not support the immediate, widespread adoption of these scores as a definitive replacement for CXR in current practice.