Improving Detection of Tuberculosis Disease Among Children

Abstract

Tuberculosis (TB) remains a major public health concern globally and is a leading infectious cause of morbidity and mortality in both adults and children. Approximately 1.3 million children and young adolescents who developed TB disease in 2023, and among those 191,000 died from TB. The risk of TB infection progressing to disease, and from disease to death, is especially high in young children. Diagnosing TB in children is difficult, with an estimated 51% children globally with TB disease not diagnosed due to the lack of a quality specimen for testing, difficulties with specimen collection and the lack of highly sensitive tests for use in children. As a result, the majority of children with TB are empirically initiated on treatment based on their symptoms alone. This dissertation aims to address specific knowledge gaps and potential areas in the patient care cascade where gaps may occur to improve detection of TB in children. In my first aim, our objective was to evaluate the clinical diagnostic accuracy of lipoarabinomannan (LAM) in urine and whole blood samples in children with clinical TB symptoms using highly sensitive electrochemiluminescent immunoassays and the most accurate known LAM antibodies (FIND28:A194-01 & S4-20:A194-01). We used data from a prospective longitudinal cohort study of predominantly HIV-negative children with clinical TB symptoms recruited from inpatient and outpatient wards at Harry Gwala Regional Hospital in KwaZulu-Natal, South Africa. Using a composite reference standard for pediatric clinical TB diagnosis, we found that neither of the two current anti-LAM antibodies did not meet the WHO target product profile criteria for a non-sputum POC test (65% sensitivity, 98% specificity) in clinically diagnosed children. More sensitive and improved LAM detection antibodies are needed if LAM is to be a viable biomarker for diagnosing TB in children without advanced HIV disease. In my second aim, we applied machine learning methods to identify clinical, biomarker and rapid diagnostic features predictive of unconfirmed TB in children in KwaZulu-Natal, South Africa. Using data from the previous study, we trained random forest models using kNN imputation and minority class oversampling strategies to predict TB status from a range of available features. Using mean decrease in accuracy and SHAP values to highlight model variable importance, we found that respiratory rate, serum C-reactive protein levels, asthmatic episodes, LAM positivity, admission for a lower respiratory tract infection and having a sick household member were offered predictive value in differentiating between unconfirmed and unlikely TB. Models incorporating inverse probability weighting to account for class imbalance prioritized medical history and exposure related predictors, while models using synthetic oversampling prioritized clinical and laboratory testing results for prediction. Including these accessible, non-sputum-based measures in pediatric TB treatment decision algorithms may improve diagnostic accuracy. In my final aim, we used an agent-based model (TBSim) to estimate the effect of introducing novel diagnostics for both adults and children on pediatric TB outcomes in uMgungundlovu district, KwaZulu-Natal, South Africa. We evaluated four separate diagnostic strategies: (1) Xpert MTB/RIF Ultra alone; (2) Xpert plus oral swabs; (3) Xpert plus FujiLAM; and (4) Xpert plus a computer assisted detection for chest X-ray(CAD-CXR) for children. We found that when used as an add-on test to Xpert Ultra, oral swabs had the greatest impact on reducing pediatric TB incidence over the simulation period, especially in children 0-4 years of age. Surprisingly, none of the interventions substantially reduced overall TB mortality compared to Xpert Ultra alone. The inclusion of TPT prevention and social household contact network in future simulations are needed to verify findings. Our findings illustrate that there are several opportunities to expand upon existing work in order to improve detection of TB disease in children. Improving detection, diagnosis and reporting of pediatric TB remains a complex challenge to address globally. Our findings highlight the need for child-centered approaches to developing novel TB diagnostics.