AI Models Show Promise for Tuberculosis Screening in Uganda and South Africa

Researchers have evaluated machine learning models for tuberculosis (TB) screening using clinical and demographic data collected from patients in Uganda and South Africa. This study represents a crucial step in applying AI to public health challenges in African contexts, specifically aiming to identify individuals who would benefit most from expensive molecular testing for TB.

The study utilized the CAGE-TB dataset, which contains information from people with presumptive TB attending community health centers in both countries. Three neural network architectures—logistic regression (LR), multilayer perceptrons (MLP), and convolutional neural networks (CNN)—were tested, incorporating greedy feature selection to enhance performance.

The findings indicate that feature selection significantly improved the area under the receiver operating characteristic (AUROC) curve for all models, with improvements ranging from 2-7%. The logistic regression model, after feature selection, achieved an AUROC of 0.8 on Ugandan data and 0.84 on South African data. While deeper networks like MLP and CNN showed promise in development, LR demonstrated more consistent performance on the held-out cohorts.

Although the LR model narrowly missed the World Health Organization's minimum sensitivity requirements, the results suggest that developing neural-network-based classifiers for TB screening is a viable approach. This research highlights the potential of AI to enhance diagnostic processes and resource allocation in healthcare systems across Africa.