Objective To establish and evaluate a mathematical model for predicting the probability of death in patients with pulmonary tuberculosis (TB) during directly observed treatment short-course (DOTS) and to provide evidences for reducing TB-related mortality． Methods Retrospective cohort analysis was used to collect the basic information, disease characteristics and treatment outcomes of 7 032 TB patients from National TB Management Information System in Tianjin from 2014 to 2017. The multivariate and unconditional Logistic regression analysis was used to select the variables, and establish a predictive model for probability of death in patients with pulmonary TB during DOTS. The goodness of fit of the predictive model for probability of death was evaluated by Hosmer-Lemeshow test, and the receiver operating charactenstic (ROC) curve was constructed to assess the performance of the prediction model. Results There were 6 711 living cases (survival group) and 321 dead cases (death group) in 7 032 TB patients during DOTS, and the mortality rate was 4.56%. Multivariate Logistic regression analysis showed that male (OR=1.922), older age (OR=1.062), retreatment of tuberculosis (OR=1.539), first sputum bacteria positive (OR=1.936), delayed seeking medication (OR=1.401) and human immunodeficiency virus (HIV) positive (OR=4.153) were independent risk factors of mortality in the TB patients. The Logistic regression equation was established as follows: Logit (P) =ln [P/(1-P)] = 0.653X1 +0.061X2 +0.431X4 +0.661X5 + 0.337X6 +1.424X9-9.191. The values of area under the ROC curve were 0.806 (95%CI：0.784-0.828). The best critical point of prediction probability was 0.054, the sensitivity was 76.36% and the specificity was 81.58%. Conclusion As a predictive model for probability of death in the TB patients during DOTS, its fitting degree and prediction efficiency of the Logistic regression model are better, and the model has better prediction value. [ABSTRACT FROM AUTHOR]