Tuberculosis Surveillance Using a Hidden Markov Model

BACKGROUND: Routinely collected data from tuberculosis surveillance system can be used to investigate and monitor the irregularities and abrupt changes of the disease incidence. We aimed at using a Hidden Markov Model in order to detect the abnormal states of pulmonary tuberculosis in Iran. METHODS: Data for this study were the weekly number of newly diagnosed cases with sputum smear-positive pulmonary tuberculosis reported between April 2005 and March 2011 throughout Iran. In order to detect the unusual states of the disease, two Hidden Markov Models were applied to the data with and without seasonal trends as baselines. Consequently, the best model was selected and compared with the results of Serfling epidemic threshold which is typically used in the surveillance of infectious diseases. RESULTS: Both adjusted R-squared and Bayesian Information Criterion (BIC) reflected better goodness-of-fit for the model with seasonal trends (0.72 and −1336.66, respectively) than the model without seasonality (0.56 and −1386.75). Moreover, according to the Serfling epidemic threshold, higher values of sensitivity and specificity suggest a higher validity for the seasonal model (0.87 and 0.94, respectively) than model without seasonality (0.73 and 0.68, respectively). CONCLUSION: A two-state Hidden Markov Model along with a seasonal trend as a function of the model parameters provides an effective warning system for the surveillance of tuberculosis.