Basic model predictors of <i>Mtb</i> infection.

Tuberculosis household contacts are at high risk of developing tuberculosis. Tuberculosis preventive therapy (TPT) is highly effective, but implementation is hindered by limited accessibility of diagnostic tests aimed at detecting Mycobacterium tuberculosis (Mtb) infection. Development of Mtb infection prediction models to guide clinical decision-making aims to overcome these challenges. We used data from 1905 tuberculosis household contacts (age ≥10 years) from Zimbabwe, Mozambique and Tanzania to develop two prediction models for Mtb infection determined by interferon-gamma release assay (IGRA) using logistic regression with backward elimination and cross-validation and converted these into a risk score. Model performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. We developed a basic model with six predictors (age, caregiver role, index case symptom duration, index HIV status, household crowding, and index GeneXpert MTB/Rif results) and a comprehensive model with eleven predictors. The basic and comprehensive risk scores showed limited predictive capability (AUROC 0.592, sensitivity 76%, specificity 35% and AUROC 0.586, sensitivity 76%, specificity 36% respectively), with considerable overlap across IGRA-positive and -negative individuals. Neither model conferred net benefit over a treat-all strategy. Overall, our results suggest that the prediction models developed in this study do not add value for guiding TPT use in high-tuberculosis burden settings. This likely reflects complex Mtb transmission dynamics at the household- and community-level, variation in individual-level susceptibility and immune response, as well as limited accuracy of IGRA testing. Improved diagnostics to determine Mtb infection status in terms of ease-of-use, accuracy, and costs are needed.