Abstract
BACKGROUND: Tuberculosis (TB) remains a significant global health challenge, necessitating reliable biomarkers for differentiation between latent tuberculosis infection (LTBI) and active tuberculosis (ATB). This study aimed to identify blood-based biomarkers differentiating LTBI from ATB through multicohort analysis of public datasets. METHODS: We systematically screened 18 datasets from the NIH Gene Expression Omnibus (GEO), ultimately including 11 cohorts comprising 2,758 patients across 8 countries/regions and 13 ethnicities. Cohorts were stratified into training (8 cohorts, n = 1,933) and validation sets (3 cohorts, n = 825) based on functional assignment. RESULTS: Through Upset analysis, LASSO (Least Absolute Shrinkage and Selection Operator), SVM-RFE (Support Vector Machine Recursive Feature Elimination), and MCL (Markov Cluster Algorithm) clustering of protein-protein interaction networks, we identified S100A12 and S100A8 as optimal biomarkers. A Naive Bayes (NB) model incorporating these two markers demonstrated robust diagnostic performance: training set AUC: median = 0.8572 (inter-quartile range 0.8002, 0.8708), validation AUC = 0.5719 (0.51645, 0.7078), and subgroup AUC = 0.8635 (0.8212, 0.8946). CONCLUSION: Our multicohort analysis established an NB-based diagnostic model utilizing S100A12/S100A8, which maintains diagnostic accuracy across diverse geographic, ethnic, and clinical variables (including HIV co-infection), highlighting its potential for clinical translation in LTBI/ATB differentiation.