1951. Optimizing implementation of artificial intelligence-informed digital X-ray screening for tuberculosis through an age- and sex-differentiated approach

Abstract Background Computer-aided detection (CAD) using artificial intelligence (AI) to analyze chest radiographs is an important tool for community tuberculosis (TB) screening in high burden countries. Most current algorithms use a universal cut-off score to select individuals for confirmatory TB testing; however, using a tailored cut-off based on client demographics (age and sex) may improve performance. Methods Community-based TB screening was conducted using portable X-ray with CAD (qXR, [Qure.ai, India]) as part of a cluster-randomized trial in Uganda. Individuals scoring above a specified threshold were offered sputum Xpert Ultra testing. This threshold was initially set at 0.5 (range: 0-1) but was later lowered to 0.2 and then 0.1 for research purposes. For clients with scores ≥ 0.1 who did not undergo Xpert testing, we used multiple imputation to infer Xpert-positive status. We assumed that those with X-ray scores < 0.1 would be Xpert-negative if tested, but considered 0.1% or 0.5% prevalence of Xpert-positive TB in sensitivity analysis. We compared the sensitivity and specificity of using a universal screening threshold of 0.5 versus tailored thresholds based on client age and sex. Results A total of 15,375 individuals were screened using AI-interpreted digital X-ray, of whom 1,748 (11.4%) had valid sputum Ultra results; 157 (9.0%) tested positive. Assuming that people with qXR scores < 0.1 would test negative on sputum Ultra, the manufacturer-recommended universal threshold of ≥ 0.5 had an estimated sensitivity of 75.8% (95%CI: 70.3-81.6) and specificity of 95.3% (95% CI 95.2-95.4). A sex-differentiated threshold (≥ 0.35 for men, ≥ 0.8 for women) had similar specificity but significantly higher sensitivity: 81.3% (95% CI 76.3-86.3). Sex-stratified thresholds also outperformed a universal threshold when assuming higher positivity among people with qXR scores < 0.1. Stratification by age did not significantly improve accuracy.Table 1:Sensitivity and specificity of different thresholds on computer-aided detection software (qXR) when using digital chest X-ray for detecting pulmonary tuberculosis.Figure 1.Distribution of CAD scores among individuals screened for TB using digital chest radiograph in Uganda (left) and the proportions of individuals with positive and trace sputum Xpert MTB/RIF Ultra results according to CAD score (right).Figure 2.Proportion of individuals with positive sputum Xpert Ultra (including trace) according to age and sex. Conclusion Using differentiated CAD score thresholds based on client sex could improve the accuracy of digital X-ray for TB screening. Future research should validate these findings in other populations and explore the value of incorporating additional client characteristics into TB screening algorithms. Disclosures All Authors: No reported disclosures