Addressing Thyroid Cancer Overdiagnosis: AI-Driven Identification, Characterization, and Outcomes of Incidental Thyroid Nodules - PROJECT SUMMARY/ABSTRACT The overdiagnosis of thyroid cancer, primarily driven by the detection of small, asymptomatic papillary thyroid cancers, imposes significant medical and financial burdens. Despite low mortality, patients often undergo unnecessary treatments, facing risks such as surgical complications and financial distress. By 2030, the annual cost of thyroid cancer care is projected to reach $3.5 billion. One key factor in thyroid nodule detection is the reporting of incidental thyroid nodules (ITNs) during imaging for non-thyroid-related concerns. ITNs appear in 10–20% of chest and neck imaging reports. With nearly 80 million computer tomographies, magnetic resonance, and other similar images performed annually in the US, millions of patients risk entering a diagnostic cascade leading to potential thyroid cancer overdiagnosis. Despite the link between increased imaging and ITNs in radiology reports, there is a substantial knowledge gap regarding the clinical outcomes of ITNs and the factors influencing their workup. Additionally, there are no standardized criteria for the appropriate evaluation of reported ITNs, hindering efforts to mitigate overdiagnosis. This project aims to reduce the unnecessary medical and financial consequences of thyroid cancer overdiagnosis. In Aim 1, we will focus on developing, externally validating, and comparing two artificial intelligence (AI) and natural language processing (NLP)-enhanced systems to identify and characterize ITNs. Using Mayo Clinic Network data, we have developed a high- performing named entity recognition (NER) system with 97% accuracy and an F1 score of 0.95 to identify ITNs and their characteristics in imaging reports. We will explore strategies to adapt large language models (LLMs) for NER, including prompting techniques, and evaluate the resilience of the models under dataset perturbations and varied report formats. Finally, the AI-NLP system will undergo external validation at the University of Florida Health Network with 4 regional sites. In Aim 2, we will deploy the NLP-enhanced AI tool for ITN identification and characterization in three large healthcare systems, including 15 regional sites representing real-world practice. We will determine the frequency of ITNs, the proportion of patients undergoing further diagnostic procedures, and the patient, clinician, practice, and ITN report factors influencing the workup and outcomes. In Aim 3, we will engage stakeholders—including patients, clinicians, and health system representatives—using a Delphi approach to develop a pathway for assessing ITN workup appropriateness. This study will validate an AI-assisted algorithm for identifying and describing ITNs across diverse imaging settings and establish guidelines to optimize ITN evaluation. The findings will support interventions to reduce low-value workups and address thyroid cancer overdiagnosis. This proposal aligns with NOT-CA-22-037 by validating an NLP-enhanced ITN identification algorithm and improving thyroid cancer overdiagnosis.
