A screening system for smear-negative pulmonary tuberculosis using artificial neural networks

Objectives: Molecular tests show low sensitivity for smear-negative pulmonary tuberculosis (PTB). A screening and risk assessment system for smear-negative PTB using artificial neural networks (ANNs) based on patient signs and symptoms is proposed. Methods: The prognostic and risk assessment models exploit a multilayer perceptron (MLP) and inspired adaptive resonance theory (iART) network. Model development considered data from 136 patients with suspected smear-negative PTB in a general hospital. Results: MLP showed higher sensitivity (100%, 95% confidence interval (CI) 78–100%) than the other techniques, such as support vector machine (SVM) linear (86%; 95% CI 60–96%), multivariate logistic regression (MLR) (79%; 95% CI 53–93%), and classification and regression tree (CART) (71%; 95% CI 45–88%). MLR showed a slightly higher specificity (85%; 95% CI 59–96%) than MLP (80%; 95% CI 54–93%), SVM linear (75%, 95% CI 49–90%), and CART (65%; 95% CI 39–84%). In terms of the area under the receiver operating characteristic curve (AUC), the MLP model exhibited a higher value (0.918, 95% CI 0.824–1.000) than the SVM linear (0.796, 95% CI 0.651–0.970) and MLR (0.782, 95% CI 0.663–0.960) models. The significant signs and symptoms identified in risk groups are coherent with clinical practice. Conclusions: In settings with a high prevalence of smear-negative PTB, the system can be useful for screening and also to aid clinical practice in expediting complementary tests for higher risk patients