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

Area of Science:

• Computational biology and bioinformatics
• Medical diagnostics and artificial intelligence

Background:

• Smear-negative pulmonary tuberculosis (PTB) presents diagnostic challenges due to low sensitivity of conventional molecular tests.
• There is a need for effective screening and risk assessment tools to improve early detection and management of smear-negative PTB.

Purpose of the Study:

• To propose and evaluate an artificial neural network (ANN)-based system for screening and risk assessment of smear-negative PTB.
• To compare the performance of different ANNs and traditional statistical models in identifying patients with smear-negative PTB.

Main Methods:

• Development of prognostic and risk assessment models using multilayer perceptron (MLP) and inspired adaptive resonance theory (iART) networks.
• Utilized data from 136 patients with suspected smear-negative PTB in a general hospital setting.
• Compared MLP performance against support vector machine (SVM) linear, multivariate logistic regression (MLR), and classification and regression tree (CART) models.

Main Results:

• The MLP model achieved the highest sensitivity (100%) and area under the receiver operating characteristic curve (AUC) (0.918).
• MLP demonstrated comparable specificity (80%) to MLR (85%), outperforming SVM linear and CART.
• Identified significant patient signs and symptoms aligned with clinical practice for risk stratification.

Conclusions:

• The proposed ANN-based system, particularly the MLP model, shows high potential for screening smear-negative PTB in high-prevalence settings.
• This system can aid clinical practice by expediting diagnostic tests for higher-risk patients, improving patient management.