A multi-country head-to-head accuracy comparison of automated chest X-ray algorithms for tuberculosis

RATIONALE: Computer-aided detection algorithms for automated chest X-ray reading have been endorsed by the World Health Organization for tuberculosis triage, but independent, multi-country assessment of current products is needed to guide implementation. OBJECTIVES: We included chest X-rays from adults who presented to outpatient facilities with at least 2 weeks of cough in India, Madagascar, the Philippines, South Africa, Tanzania, Uganda, and Vietnam. METHODS: We calculated and compared the accuracy overall and by country and key groups for 7 computer-aided detection algorithms: CAD4TB, qXR, INSIGHT CXR, DrAid, Genki, InferRead, and Radify. We determined if any computer-aided detection product could achieve the minimum target accuracy for a tuberculosis triage test (≥ 90% sensitivity and ≥ 70% specificity). RESULTS: Of 3901 individuals included, the median age was 41 years (IQR, 29-54 years), 12.9% were people living with HIV, 8.2% were living with diabetes, and 21.2% had a prior history of tuberculosis. Specificity ranged from 30.9% to 73.5% at 90% sensitivity. CAD4TB achieved the highest specificity at 90% sensitivity (73.5% specific [95% CI, 71.9%-75.1%]), although qXR and INSIGHT CXR also achieved the target 70% specificity. There was heterogeneity by country and subgroup that improved with population-specific thresholds, except for people living with HIV, 50 years and older, or with a history of tuberculosis. CONCLUSIONS: Multiple computer-aided detection algorithms achieved the minimum target accuracy for a tuberculosis triage test among symptomatic individuals with cough. Further efforts are needed to integrate computer-aided detection into routine tuberculosis case detection programs in high-burden communities.