Artificial intelligence demonstrates non-inferiority to radiologists at tuberculosis detection within a population with a high-disease burden
1. Kazemzadeh and colleagues evaluated the performance of AI systems that detect TB and chest X-ray abnormalities against that of radiologists.
2. AI demonstrated non-inferiority to the radiologists for active TB detection and the potential to effectively detect other chest X-ray abnormalities.
Evidence Rating Level: 2 (Good)
Study Rundown: A Chest X-ray (CXR) can improve early tuberculosis (TB) detection, but in low-to-middle-income countries, there is a lack of human experts to interpret imaging results. Artificial intelligence (AI) has the potential to automate CXR readings. Kazemzadeh and colleagues evaluated two AI systems, one detecting TB (TB AI) and one detecting general CXR abnormalities (abnormality AI) in patients from an endemic area. Sputum samples were collected from participants to confirm TB diagnosis, and a reference standard for CXR abnormality was developed by experienced radiologists. AI’s sensitivity and specificity were compared against the World Health Organization (WHO) recommendations and radiologists who were instructed to read CXR without clinical information. The TB AI demonstrated higher sensitivity compared to radiologists but fell short of the WHO-recommended sensitivity. For specificity, the TB AI was inferior to that of radiologists but did reach WHO’s specificity target. The abnormality AI reached predetermined sensitivity and specificity targets. This study demonstrated TB AI’s non-inferiority to radiologists for active TB detection within a population with a high TB burden. Meanwhile, the abnormality AI could detect other CXR abnormalities in the same population.
Click here to read the study in NEJM AI
Relevant Reading: Real-world testing of an artificial intelligence algorithm for the analysis of chest X-rays in primary care settings
In-Depth [prospective cohort]: 1827 adults from Zambia were included in the study if they exhibited cardinal TB symptoms, were in household contact with a TB patient, and had newly diagnosed HIV status. Sputum culture was used to confirm TB status, and the reference standard for CXR abnormality was developed by three radiologists based in India with a minimum of 8 years’ experience. Ten India-based radiologists (mean experiences of experience 9.1±2.0, 6-12) assessed CXR for TB and were instructed to read for sensitivity and that the CXR was from a TB endemic region but were blinded to clinical information. The sensitivity and specificity of the two AI systems were compared against radiologist performance and the WHO recommendation for TB detection (90% sensitivity and 70% specificity for TB). For CXR abnormalities, the endpoint was 90% sensitivity and 50% specificity. Radiologists achieved a mean sensitivity of 76% (95% confidence interval [CI], 71-83) and a mean specificity of 82% (95% CI, 77-85). The TB AI achieved a higher sensitivity of 87% (95% CI, 82-92), but a lower specificity of 70% (95% CI, 67-72). For CXR abnormalities, the abnormality AI achieved a sensitivity of 97% (95% CI, 95-98) and a specificity of 79% (95% CI, 77-81), both of which met the pre-determined endpoint. The study’s limitations included consistent artifacts on images produced by the CXR machine, which may have affected AI performance and radiologists’ inability to reach WHO targets. Overall, this study provided encouraging non-inferiority evidence for AI-based TB screening in endemic areas compared to radiologist interpretations.
Image: PD
©2025 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
