Real-World Experience With Artificial Intelligence-Enabled Natural Language Processing for Identification and Management of Incidental Lung Nodules.
Authors
Affiliations (7)
Affiliations (7)
- Hematology/Oncology Fellowship, Department of Graduate Medical Education, Orlando Health Cancer Institute, Orlando, FL, USA.
- Oncology Quality, Orlando Health Cancer Institute, Orlando, FL, USA.
- Internal Medicine Residency, Department of Graduate Medical Education, Orlando Regional Medical Center, Orlando, FL, USA.
- Oncology Service Line, Orlando Health Cancer Institute, Orlando, FL, USA.
- Section of Thoracic Surgery, Division of Thoracic Oncology, Department of Surgical Oncology, Orlando Health Cancer Institute, Orlando, FL, USA.
- Section of Interventional Pulmonology, Division of Thoracic Oncology, Department of Surgical Oncology, Orlando Health Cancer Institute, Orlando, FL, USA.
- Section of Thoracic, Head & Neck Oncology, Division of Thoracic Oncology, Department of Medical Oncology, Orlando Health Cancer Institute, Orlando, FL, USA.
Abstract
BackgroundIncidental lung nodules (ILNs) are common, but adherence to guideline-recommended follow-up is variable, which may contribute to delayed diagnosis of malignancy. Artificial intelligence (AI) platforms using natural language processing (NLP) may improve identification, but their impact on clinical management and outcomes remains uncertain.ObjectivesThis study describes ILN identification, follow-up management, and timeliness of cancer treatment after implementation of an AI-enabled NLP platform at a large health system.DesignSingle-center retrospective observational study.MethodsAdults with ILNs identified on imaging between April 2023 and May 2024 were included. The pre-AI cohort was identified using diagnosis codes, while the post-AI cohort was identified using an NLP-based AI platform. Because these methods captured different populations, no formal comparative testing was performed, and outcomes were summarized descriptively.ResultsA total of 1,098 patients with ILNs were identified: 228 from 6,828 pre-AI imaging studies (3.3%) and 870 from 9,902 post-AI imaging studies (8.8%). Most pre-AI ILNs were identified on outpatient imaging (74%), while post-AI ILNs were identified in inpatient (35%), emergency department (26%), and outpatient (39%) settings. A follow-up recommendation was documented for 195 of 228 pre-AI patients (85.5%) and 563 of 870 post-AI patients (64.7%). Among these patients, completion was documented for 167 of 195 (85.6%) and 414 of 563 (73.5%), respectively. Malignancy was documented in 28 of 228 pre-AI patients (12.3%) and 15 of 870 post-AI patients (1.7%). Median imaging-to-biopsy time was 78 days pre-AI and 120 days post-AI; biopsy-to-treatment and imaging-to-treatment intervals were similar.ConclusionThe post-AI cohort included more identified ILNs and a greater proportion from acute-care settings than the historical pre-AI cohort. Because identification methods were not comparable, observed differences in documented malignancy diagnosis and treatment timing cannot be attributed to the platform. Effective implementation requires defined clinical ownership and follow-up infrastructure.