Performance of five dynamic models in predicting tuberculosis incidence in three prisons in Thailand.

This study examined the ability of the following five dynamic models for predicting pulmonary tuberculosis (PTB) incidence in a prison setting: the Wells-Riley equation, two Rudnick & Milton-proposed models based on air changes per hour and liters per second per person, the Issarow et al. model, and the applied susceptible-exposed-infected-recovered (SEIR) tuberculosis (TB) transmission model. This 1-year prospective cohort study employed 985 cells from three Thai prisons (one prison with 652 cells as the in-sample, and two prisons with 333 cells as the out-of-sample). The baseline risk of TB transmission for each cell was assessed using the five dynamic models, and the future PTB incidence was calculated as the number of new PTB cases per cell and the number of new PTB cases per 1,000 person-years (incidence rate). The performance of the dynamic models was assessed by a four-step standard assessment procedure (including model specification tests, in-sample model fitting, internal validation, and external validation) based on the Negative Binomial Regression model. A 1% increase in baseline TB transmission probability was associated with a 3%-7% increase in future PTB incidence rate, depending on the dynamic model. The Wells-Riley model exhibited the best performance in terms of both internal and external validity. Poor goodness-of-fit was observed in all dynamic models (chi-squared goodness-of-fit tests of 70.75-305.1, 8 degrees of freedom, p < .001). In conclusion, the Wells-Riley model was the most appropriate dynamic model, especially for large-scale investigations, due to its fewer parameter requirements. Further research is needed to confirm our findings and gather more data to improve these dynamic models.