Towards probabilistic decision support in public health practice: Predicting recent transmission of tuberculosis from patient attributes

Objective: Investigating the contacts of a newly diagnosed tuberculosis (TB) case to prevent TB transmission is a core public health activity. In the context of limited resources, it is often necessary to prioritize investigation when multiple cases are reported. Public health personnel currently prioritize contact investigation intuitively based on past experience. Decision-support software using patient attributes to predict the probability of a TB case being involved in recent transmission could aid in this prioritization, but a prediction model is needed to drive such software. Methods: We developed a logistic regression model using the clinical and demographic information of TB cases reported to Montreal Public Health between 1997 and 2007. The reference standard for transmission was DNA fingerprint analysis. We measured the predictive performance, in terms of sensitivity, specificity, negative predictive value, positive predictive value, the Receiver Operating Characteristic (ROC) curve and the Area Under the ROC (AUC). Results: Among 1552 TB cases enrolled in the study, 314 (20.2%) were involved in recent transmission. The AUC of the model was 0.65 (95% confidence interval: 0.61-0.68), which is significantly better than random prediction. The maximized values of sensitivity and specificity on the ROC were 0.53 and 0.67, respectively. Conclusions: The characteristics of a TB patient reported to public health can be used to predict whether the newly diagnosed case is associated with recent transmission as opposed to reactivation of latent infection. (C) 2014 Elsevier Inc. All rights reserved.