Published
Union Conference on Lung Health (2024)
Authors
J Lecciones1, JP Ubalde1, MR Santiago2, N Marquez2, S Guirgis2, L Stevens3
Affiliations
1Tropical Disease Foundation, Philippines,
2Family Health International (FHI) 360, Philippines,
3FHI 360, Asia Pacific Regional Office, Thailand
*to be published soon.
Summary
Chest X-ray equipped with computer-aided detection powered by artificial intelligence software (CAD-AI) is a key strategy for TB case finding in the Philippines. Evaluating CAD-AI performance during active case finding in communities can inform its uptake and expansion in TB mass screening programs.
Background
The use of chest X-ray (CXR) equipped with computer-aided detection powered by artificial intelligence software (CAD-AI) in TB screening was introduced in 2019. However, expanding the use of CAD-AI in active case finding (ACF) activities is quite challenging due to perceived uncertainty of its performance and usefulness as a relatively new tool.
Intervention
ACF using symptoms screening and CXR with CAD-AI was implemented in two highly urbanized cities in Metro Manila targeting urban poor communities. CAD-AI provided a threshold score (0-100%) for each radiograph with a pre-specified threshold of 30% indicating presumptive TB. Radiographs were read by experienced radiologists who made dichotomous decisions on whether findings were suggestive of TB. Individuals with presumptive TB either by symptoms, CAD-AI or human readers were advised to submit sputum specimens for Xpert MTB/Rif testing. The performance of CAD-AI and radiologists in TB detection were compared with reference to Xpert MTB/Rif test results, using McNemar’s test with 95% confidence intervals (CIs).
Results
Between August and October 2023, 20,042 individuals were screened for TB, of whom 3,267 were tested with Xpert MTB-Rif, and 311 bacteriologically confirmed (BC)-TB cases were identified. The sensitivity of CAD-AI and radiologists in detecting TB was the same at 95% (p=0.85), while specificity of CAD-AI was significantly higher than the radiologists (53%, 95% CI [51%, 55%] for CAD-AI vs. 49%, 95% CI [47%, 51%] for radiologists, p<0.001). When combining TB-positive predictions from CAD-AI and radiologists, the sensitivity significantly increased to 99% (p<0.001), and only 3 of 311 BC-TB were missed. However, in this case, specificity decreased compared to either CAD-AI or radiologists (39%, CI [37%, 41%], p<0.001).
Conclusion
The results suggest that CAD-AI has the potential to become an important tool in promptly and accurately identifying presumptive TB during mass TB screening programs. Hence, increasing uptake and expanding use of CAD-AI should be considered.