Computer-aided detection pragmatic threshold calibration in TB active case finding: A mixed-methods study in Tondo, Manila
PLOS Global Public Health, 2026
Abstract
Introduction: In high-burden TB settings, managing the operational balance between screening sensitivity and referral workload is a critical challenge. This study evaluated a pragmatic framework for calibrating AI-based computer-aided detection (CAD) thresholds within a TB active case-finding (ACF) programme in Tondo, Manila — one of the Philippines' most densely populated urban communities — assessing both its feasibility and acceptability among frontline healthcare workers and stakeholders.
Methods: CAD4TB v.7 (Delft Imaging, Netherlands) was integrated into ACF activities targeting individuals aged 15 years and older from November 2022. Referral for sputum collection and Xpert molecular testing was triggered by CAD scores exceeding a defined threshold. A mixed-methods design was employed: retrospective analysis of pre-CAD radiologist data (May–November 2022) to establish an initial threshold; prospective monitoring (November 2022–October 2023) to guide iterative threshold adjustments; and semi-structured interviews with healthcare workers and operational stakeholders (January–April 2023). Threshold decisions were informed by a combination of quantitative performance data, user feedback, and on-the-ground operational observations.
Results: An initial CAD threshold of 25 was selected to replicate the 35% referral rate previously achieved by onsite radiologists. Within two weeks, the referral rate climbed to 37.5%, prompting an upward adjustment to a threshold of 28. Eight months into the programme, referral stood at 32.6% with a screening yield of 4.9% (Xpert-positive cases among all screened individuals). Following further stakeholder consultation and data review, the threshold was raised again to 32, after which the referral rate fell below 25% and yield settled at 4.2%. Across the full study period (October 2022–October 2023), 14,739 individuals were screened, producing an overall screening yield of 4.3% and identifying 633 confirmed TB cases.
Conclusion: Iterative, data-driven calibration of CAD thresholds — anchored to referral rate targets, screening yield, and operational capacity — can meaningfully improve ACF efficiency while building user confidence in AI tools. Real-time monitoring combined with structured stakeholder input enabled the programme to balance detection sensitivity against the practical constraints of a mobile screening service in an underserved population. The authors propose this framework as a transferable model for other high-burden settings, though further validation across diverse populations and exploration of subgroup-specific thresholds are recommended.