Abstract Rationale Lung nodules are frequently found incidentally on imaging, which necessitates effective management strategies. The Fleischner Society guidelines outline management for nodules, but adherence is variable, and missed follow-up may delay cancer diagnosis. Traditional identification methods rely on International Classification of Diseases (ICD) based coding, which captures patients already recognized by providers. Artificial intelligence (AI) may improve detection and management of incidental lung nodules (ILNs). In contrast to traditional methods, AI-driven natural language processing (NLP) platforms can extract nodule data from radiology reports, increasing detection and supporting management. Methods This single-center retrospective study evaluated adults who underwent imaging between April 2023 and May 2024. Two cohorts were analyzed: a pre-AI period (April to September 2023) when nodules were identified by ICD codes, and a post-AI period (October to May 2024) when nodules were identified using an NLP-based AI platform (Thynk Health, Lexington, Kentucky). Variables included demographics, imaging location and type, and management. In the post-AI period, a lung nodule navigator was introduced to tracked findings, communicate with patients and providers, and coordinate follow-up as able. Results Among 1098 patients, the platform increased ILN identification (228 pre-AI vs 870 post-AI). Pre-AI detections occurred mainly in outpatient imaging (74%), while post-AI expanded to inpatient (35%) and emergency department settings (26%). Pre-AI management included referral to pulmonology (23%), thoracic surgery or interventional pulmonology (12%), oncology (4%), serial imaging (56%) and no follow-up (5%). Post-AI management included referral to pulmonology (7%), thoracic surgery or interventional pulmonology (6%), oncology (1%), serial imaging (50%) and no follow-up (35%). Median time from imaging to biopsy was longer post-AI (120 vs 78 days), while imaging to treatment intervals were similar (129 vs 109 days). Malignant ILNs represented 12% of pre-AI and 1.7% of post-AI. Conclusions Real-world use of an AI-driven NLP platform increased identification of ILN across care settings but revealed challenges in follow-up and management. The pre-AI cohort, identified by ICD codes, was more likely to have findings actively managed, contributing to shorter imaging to biopsy times. Despite the increased identification of ILNs, the lack of increase in cancer detection or earlier treatment suggests that identification alone may not result in improved clinical outcomes. Follow-up efforts were limited by systemic and insurance barriers that restricted access to specialty care. In response to rising volumes, our institution expanded navigator roles with a multidisciplinary protocol for tracking and communication to improve timely evaluation and outcomes. This abstract is funded by: None
Ploucher et al. (Fri,) studied this question.