Bacterial load slopes represent biomarkers of tuberculosis therapy success, failure, and relapse

There is an urgent need to discover biomarkers that are predictive of long-term TB treatment outcomes, since treatment is expense and prolonged to document relapse. We used mathematical modeling and machine learning to characterize a predictive biomarker for TB treatment outcomes. We computed bacterial kill rates, γ f for fast- and γ s for slow/non-replicating bacteria, using patient sputum data to determine treatment duration by computing time-to-extinction of all bacterial subpopulations. We then derived a γ s- slope-based rule using first 8 weeks sputum data, that demonstrated a sensitivity of 92% and a specificity of 89% at predicting relapse-free cure for 2, 3, 4, and 6 months TB regimens. In comparison, current methods (two-month sputum culture conversion and the Extended-EBA) methods performed poorly, with sensitivities less than 34%. These biomarkers will accelerate evaluation of novel TB regimens, aid better clinical trial designs and will allow personalization of therapy duration in routine treatment programs. Magombedze et al. propose a new method combining mathematical modeling and machine learning to derive early (within 2 months) effective predictive biomarkers from bacterial load slopes for tuberculosis long term treatment outcomes, thereby accurately predicting treatment failure and success.