Screening for active pulmonary tuberculosis: Development and applicability of artificial neural network models

Abstract

Tuberculosis (TB) remains a significant public health challenge, motivated by the diversity of healthcare epidemiological settings, as other factors. Cost-effective screening has substantial importance for TB control, demanding new diagnostic tools. This paper proposes a decision support tool (DST) for screening pulmonary TB (PTB) patients at a secondary clinic. The DST is composed of an adaptive resonance model (iART) for risk group identification (low, medium and high) and a multilayer perceptron (MLP) neural network for classifying patients as active or inactive PTB. Our tool attains an overall sensitivity (SE) and specificity (SP) of 92% (95% CI; 79-97) and 58% (95% CI; 47-68), respectively. SE values for smear-positive and smear-negative patients are 96% (95% CI; 80-99) and 82% (95% CI; 52-95), as well as higher than 83% (95% CI; 43-97) in low and high-risk cases. Even in scenarios with prevalence up to 20%, negative predictive values superior to 95% are obtained. The proposed DST provides a quick and low-cost pretest for presumptive PTB patients, which is useful to guide confirmatory testing and patient management, especially in settings with limited resources in low and middle-incoming countries.

Keywords: Decision support systems; Diagnosis; Neural network models; Tuberculosis.