Development and validation of a prediction model for active tuberculosis case finding among HIV-negative/unknown populations

Abstract

A prediction model of prevalent pulmonary tuberculosis (TB) in HIV negative/unknown individuals was developed to assist systematic screening. Data from a large TB screening trial were used. A multivariable logistic regression model was developed in the South African (SA) training dataset, using TB symptoms and risk factors as predictors. The model was converted into a scoring system for risk stratification and was evaluated in separate SA and Zambian validation datasets. The number of TB cases were 355, 176, and 107 in the SA training, SA validation, and Zambian validation datasets respectively. The area under curve (AUC) of the scoring system was 0·68 (95% CI 0·64-0·72) in the SA validation set, compared to prolonged cough (0·58, 95% CI 0·54-0·62) and any TB symptoms (0·6, 95% CI 0·56–0·64). In the Zambian dataset the AUC of the scoring system was 0·66 (95% CI 0·60–0·72). In the cost-effectiveness analysis, the scoring system dominated the conventional strategies. The cost per TB case detected ranged from 429 to 1,848 USD in the SA validation set and from 171 to 10,518 USD in the Zambian dataset. The scoring system may help targeted TB case finding under budget constraints.