Marina E Robson Chase, Jacob Hubbuch, Colleen A McMullen, Daniel Davenport, Omar El Kawkgi, Nora Habib, Alexandria Willis, Rachel Zalla, Danielle S Walsh, Oliver J Fackelmayer, Cortney Y Lee, William B Inabnet
Large language models that are not specifically trained for differentiated thyroid cancer management demonstrate variability in recommending radioactive iodine therapy. However, these platforms may complement human recommendations for radioactive iodine administration, with promise toward aligning radioactive iodine use with consensus guidelines.
BACKGROUND: The decision to administer postoperative radioactive iodine therapy in differentiated thyroid cancer is evolving. Given the growing utility of large language models in medical decision-making, we hypothesize that artificial intelligence models can inform recommendations regarding postoperative radioactive iodine therapy.
METHODS: After institutional review board approval, demographic, operative, and pathologic details from 100 adult patients with differentiated thyroid cancer who underwent total thyroidectomy were retrospectively extracted. A comprehensive literature review was conducted to model an appropriate radioactive iodine consensus-informed estimate for each American Thyroid Association risk category. Four artificial intelligence platforms were queried for recommendations on the utility of postoperative radioactive iodine for each patient. The physician prescribed radioactive iodine treatment (human recommendation) and results of the artificial intelligence queries (artificial intelligence recommendation) were compared with the estimated radioactive iodine rates using binomial exact tests.
RESULTS: Of the 100 patients reviewed, the median age was 48 years, and 81% were female. The 2015 American Thyroid Association risk stratification resulted in 59 low-, 25 intermediate-, and 16 high-risk patients. Physicians and Google Gemini recommended radioactive iodine in 24% and 19% of low-risk patients, respectively, which was statistically greater than the consensus-informed estimate of 10%. In intermediate-risk patients, both physicians (92%) and Doximity GPT (76%) exceeded the radioactive iodine consensus-informed estimate of 50%. In high-risk patients, Open Evidence recommended radioactive iodine at a significantly lower rate (69%) than all comparators.
CONCLUSION: Large language models that are not specifically trained for differentiated thyroid cancer management demonstrate variability in recommending radioactive iodine therapy. However, these platforms may complement human recommendations for radioactive iodine administration, with promise toward aligning radioactive iodine use with consensus guidelines.