Cost-effectiveness evaluation of artificial intelligence-assisted chest radiograph interpretation for tuberculosis screening in rural health units in the Philippines

Tuberculosis remains a major public health burden in the Philippines, where diagnostic delays are amplified by limited radiology capacity in rural health units (RHUs) and geographically isolated and disadvantaged areas (GIDAs). Computer-aided diagnosis (CAD) using artificial intelligence (AI)-assisted chest radiograph interpretation may shorten the screening pathway and reduce reliance on scarce specialist readers. However, its economic value for RHU-based tuberculosis screening has not been fully evaluated. We developed a decision-tree cost-effectiveness model in Microsoft Excel 365 to compare AI-assisted chest radiograph interpretation with conventional manual radiologist or teleradiology interpretation among a theoretical annual cohort of 1,000 presumptive tuberculosis patients in RHU settings. The model adopted a provider perspective over a 5-year horizon, used 2026 Philippine pesos, and applied a base-case discount rate of 5%, with 3% and 7% explored in scenario analyses. Diagnostic performance inputs were derived from published CAD evaluations, while cost inputs were based on local program costing assumptions, public-sector salary and diagnostic cost estimates, and supplier quotations where available. Outcomes were expressed as additional true tuberculosis cases correctly detected, discounted incremental costs, and the incremental cost-effectiveness ratio (ICER). One-way, multi-way, and diagnostic-performance scenario analyses were conducted. In the base case, AI-assisted interpretation detected 90.22 true tuberculosis cases per 1,000 screened compared with 88.72 under manual interpretation, while reducing false-positive radiographs from 453.51 to 286.11. Annual costs were Php 877,330 for AI-assisted interpretation and Php 1,142,230 for manual interpretation. Over 5 years, the discounted incremental cost was Php − 946,878 and the discounted incremental effectiveness was 6.49 additional correctly detected tuberculosis cases, yielding a dominant ICER of Php − 145,803 per additional case detected. The base-case conclusion remained cost-saving under most deterministic sensitivity analyses. When a lower manual/teleradiology reading fee of Php 300 per scan or Philippine qXR pseudo-accuracy estimates were applied, AI-assisted interpretation was no longer cost-saving but remained more effective, emphasizing the importance of local calibration and budget-impact assessment. This decision-analytic model suggests that AI-assisted chest radiograph interpretation may be a cost-saving and more effective tuberculosis screening strategy for RHU settings under base-case assumptions, largely by reducing unnecessary confirmatory GeneXpert testing. The findings support pilot implementation and targeted evaluation in high-burden, underserved settings; however, national scale-up should be accompanied by local diagnostic validation, implementation monitoring, and budget-impact analysis.