Fig. 2

Machine learning models (biosignatures) discriminating between progressors and controls. Panels show receiver–operator characteristic (ROC) curves. The three panels correspond to the three models tested: a, model Total which was generated using all training set samples; b, model Total/Baseline which was generated using only BL training set samples. Model evaluation was stratified by time to TB diagnosis: all, evaluation on all test set samples; proximate, evaluation on test set samples collected < 5 months before TB diagnosis; distal: ≥ 5 months. c Results of enrichment test on metabolites ordered by their importance in the Total and Total/Baseline models. The metabolite sets correspond to biochemical groups and clusters of metabolites identified previously in TB patients. Color intensity corresponds to p-value, and symbol size corresponds to the strength of the enrichment. P-values were corrected for multiple testing, and AUC was used as a measure of effect size