Artificial neural network (ANN) for prediction of pulmonary tuberculosis in hospitalized patients

Tuberculosis (TB) endangers all individuals, irrespective of country of origin. Hospitals have an established relationship with transmission of TB, and a delay in detection of TB can lead to the spread of disease. Rio de Janeiro has a high burden of TB and up to 1/3 of its cases are diagnosed in hospitals. Rapid diagnosis and respiratory isolation (RI) are essential to reduce nosocomial transmission. Unfortunately, the costs of isolation rooms (IR) are high, so optimization of its use is necessary. Clinical models have been described to reduce RI, without an increase in the risk of transmission. We developed an ANN to predict pulmonary TB in a high complexity hospital in Rio de Janeiro. The network is a multilayer perceptron (MLP) network, with clinical and radiological data collected from a cohort of 290 patients admitted to IR due to suspicion of TB. The ANN has one hidden layer, and was trained with the resilient back propagation algorithm and has the hyperbolic tangent as the activation function. The training group consisted of 80% of the patients. The remaining were used for validation. The stop criterion was the SP product. The neural model included 21 variables and has a sensitivity of 100%, specificity of 99.3%, positive and negative predictive values of 98.3% and 100%, respectively. The variable with the greatest discriminative value is chest X-ray. HIV status was a cofounder, so the final model did not include it. In conclusion, the ANN developed had a high accuracy for pulmonary TB diagnosis and could be used as a tool to optimize use of IR in hospitals. However, this model should be prospectively evaluated before routine use in distinct epidemiological scenarios.