Computer-aided detection pragmatic threshold calibration in TB active case finding: A mixed-methods study in Tondo, Manila

We developed a pragmatic approach to calibrating computer-aided detection (CAD) thresholds in a tuberculosis (TB) active case finding (ACF) project in the Philippines, assessing its feasibility and acceptability in a high-burden urban setting. CAD4TB v.7 (Delft Imaging, Netherlands) software was integrated into ACF activities targeting individuals aged ≥15 years in Tondo, Manila, from November 2022. Referrals for sputum collection and Xpert testing were based on CAD scores above a threshold. We used a mixed-methods approach: 1) retrospective analysis of pre-CAD data (May-November 2022) to determine the initial threshold; 2) prospective monitoring (November 2022-October 2023) to adjust the threshold; 3) semi-structured interviews with healthcare workers and stakeholders (January-April 2023). Threshold selection combined quantitative data, user experiences, and operational observations. The initial CAD threshold of 25 matched the 35% referral rate of onsite radiologists. After two weeks, referral rose to 37.5%, prompting an increase to 28. Eight months later, referral was 32.6% with a 4.9% screening yield (Xpert positive among screened individuals). Following data and stakeholders’ input, the threshold was further increased to 32, after which referral dropped below 25% and the yield was 4.2%. Overall, 14 739 individuals were screened from October 2022 to October 2023, with a total screening yield of 4.3%, resulting in 633 TB cases. CAD threshold calibration, informed by referral rate targets, screening yield, and operational constraints can improve ACF efficiency and user confidence. Adjustments informed by real-time monitoring and stakeholder feedback helped balance sensitivity and operational feasibility in a mobile, underserved population. This framework supports tailoring CAD deployment to local epidemiology, software versioning, and system capacity. It is applicable to other high-burden settings, although further validation across diverse populations and assessment of subgroup-specific thresholds are warranted.