Prospective Multi-Site Validation of AI to Detect Tuberculosis and Chest X-Ray Abnormalities.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Artificial intelligence (AI) holds potential for interpreting chest X-rays (CXRs) to support accessible triage for active pulmonary tuberculosis (TB).
• This is particularly relevant in resource-constrained settings where TB prevalence is high.

Purpose of the Study:

• To evaluate the performance of two cloud-based CXR AI systems for TB detection and general abnormality detection.
• To compare AI performance against radiologists and World Health Organization (WHO) targets in a population with high TB and HIV burden.

Main Methods:

• Two AI systems were evaluated: one for TB detection (TB AI) and another for general CXR abnormalities (abnormality AI).
• Performance was assessed by comparing AI results to radiologist interpretations and WHO targets (90% sensitivity, 70% specificity) in 1978 adults with TB symptoms, TB contacts, or new HIV diagnoses.
• AI performance was analyzed for noninferiority against radiologists and predefined sensitivity/specificity targets.

Main Results:

• The TB AI demonstrated 87% sensitivity and 70% specificity at a high-sensitivity threshold, and 78% sensitivity and 82% specificity at a balanced threshold.
• The TB AI was noninferior to radiologists' mean sensitivity (76%) at the high-sensitivity threshold but not specificity (82%).
• The abnormality AI achieved 97% sensitivity and 79% specificity, meeting its prespecified targets.

Conclusions:

• CXR AI demonstrated noninferiority to radiologists for active pulmonary TB triaging in a high TB and HIV burden population.
• Neither the TB AI nor the radiologists met WHO sensitivity recommendations for the study population.
• AI is a viable tool for detecting CXR abnormalities beyond TB in this demographic.