Closing the diagnostic gap: A three-step AI safety-net workflow to capture missed lung nodules on chest radiographs in routine clinical practice.

3

Background: Incidentally encountered lung nodules on chest radiographs (CXRs) are frequently overlooked in routine clinical practice. A Thai institutional study demonstrated that up to 79% of lung cancers visible in retrospect on CXRs were missed by both radiologists and clinicians (Chutivanidchayakul et al., Clin Imaging 2023). Even when AI systems flag suspicious nodules, clinician non-concordance remains a critical barrier. We implemented a three-step safety-net workflow integrating AI detection, systematic monitoring, and radiologist second-look review through a dedicated lung nodule clinic to close this diagnostic gap. Methods: Over 1 year, all CXRs obtained for any clinical indication at Phrapokklao Hospital and its spoke network in Chanthaburi, Thailand, were analyzed by qXR (Qure.ai) for lung nodule detection. The workflow comprised: (1) AI-flagged detection of actionable high-risk lung nodules (HLNMs), (2) centralized monitoring to capture all flagged cases regardless of treating physician response, and (3) radiologist second-look review to validate findings and refer confirmed cases for diagnostic CT and biopsy. Lung cancers were histologically confirmed and staged per AJCC 8th edition. Results: Of 107,238 CXRs processed, qXR identified 691 actionable HLNMs (0.64%). Through the safety-net pathway, 382 (55.3%) underwent diagnostic CT and 124 (32.5%) proceeded to biopsy. Forty-seven lung cancers were confirmed (1 per 2,282 CXRs), plus 2 non-lung metastatic cancers. Among 33 patients with complete staging: stage I (n=3), stage II (n=11), stage IIIA–C (n=9), and stage IV (n=10). Overall, 23/47 (48.9%) were early-stage (I–III). Conclusions: AI detection alone is insufficient when clinicians fail to act on flagged findings. Our three-step safety-net model effectively bridges the gap between AI identification and clinical action, capturing lung cancers that would otherwise be missed. Nearly half were early-stage, demonstrating meaningful diagnostic acceleration. This approach addresses an underrecognized failure point in AI-augmented workflows and offers a scalable model for reducing missed lung cancer in routine practice.