Artificial Intelligence-Enabled Analysis of Radiology Reports: Epidemiology and Outcomes of Incidental Thyroid Findings

The study built and deployed a transformer-based natural language processing pipeline that identified incidental thyroid findings and extracted nodule characteristics from imaging reports across multiple modalities and body regions. Prevalence, features, and consequences of incidental thyroid findings were previously undefined; this pipeline converts unstructured radiology reports into scalable epidemiologic data. Retrospective cohort of 115,683 patients across multiple Mayo Clinic sites, with a large and multi-institutional sample.

Among 115,683 patients, 9,077 (7.8%) had an incidental thyroid finding, 92.9% of which were nodular; findings were more likely in women, older adults, those with higher BMI, and imaging ordered by specialties other than Emergency Medicine. First to quantify overall prevalence and demographic and imaging-related factors of incidental thyroid findings in a large multi-site cohort. Retrospective cohort design with logistic regression to identify associated factors, sample size exceeding one hundred thousand.

Incidental thyroid findings were more common on neck CT, PET, and nuclear medicine scans compared with chest CT; nodule characteristics were poorly documented, with size reported in only 44% and other features in fewer than 15%. Reveals differences in detection rates across imaging modalities and quantifies the extent of missing nodule characterization in radiology reports. Descriptive statistics from the same cohort covering multiple imaging modalities and body regions.

Compared with patients without incidental thyroid findings, those with such findings had higher odds of subsequent thyroid nodule diagnosis (OR 45, 95% CI 41.1-49.3), biopsy (OR 46.8, 95% CI 39.0-56.2), thyroidectomy (OR 55.8, 95% CI 31.3-99.3), and thyroid cancer diagnosis (OR 61.7, 95% CI 38.6-98.5); most cancers were papillary (88.5%), and cancers detected after incidental findings were larger (2 cm, SD 1.4) versus no incidental finding (1.3 cm, SD 0.8). Links incidental thyroid findings to a cascade of downstream clinical interventions and cancer diagnoses, highlighting their role in overdiagnosis. Odds ratio estimates from a large cohort, with wide confidence intervals for some outcomes (e.g., thyroidectomy and cancer diagnosis), indicating limited precision.

The study applies to adults without prior thyroid disease undergoing thyroid-capturing imaging, based on retrospective data from Mayo Clinic sites from 2017 to 2023; its natural language processing pipeline and epidemiologic findings can inform similar report analyses and follow-up strategy evaluations in other health systems, but local reporting practices and population characteristics should be considered.

As a fast parse, this summary lacks figures and supplementary materials, so specific performance metrics of the natural language processing pipeline (e.g., accuracy, recall) and validation details of nodule characteristic extraction cannot be assessed; additionally, the wide confidence intervals for thyroidectomy and cancer diagnosis odds ratios indicate limited precision, and future work should validate in more diverse populations and evaluate the real-world effects of standardized reporting and selective follow-up.