Clinical predictors of I-131 therapy failure in differentiated thyroid cancer by machine learning: A single-center experience.

Abstract

Well-differentiated thyroid carcinoma is predominantly a slow-growing malignancy, amendable to treatment, and has an excellent prognosis following thyroidectomy and radioiodine (RAI) therapy. However, patients who fail the initial RAI treatment attempt may require repeated RAI or other treatments and with this, comes an associated impact on patient quality of life. Therefore, the anticipation of patients in whom there is a higher risk of RAI failure may help in patient risk stratification and subsequent management. We conducted a retrospective review to determine the factors associated with initial RAI therapy failure in well-differentiated thyroid cancer patients. Using scikit-learn from Python, we implemented a machine-learning algorithm to determine the clinical patient factors associated with a higher likelihood of treatment resistance. We found that clinical factors such as tumor focality (P = 0.026) and lymph node invasion at surgical resection (P = 0.0135) were significantly associated with initial treatment failure following RAI. Elevated serum thyroglobulin (Tg) and Tg antibody levels following surgery but before RAI were also associated with treatment resistance (P < 0.0001 and P = 0.011 respectively). Less expected factors such as decreased time from surgery to RAI were also associated with treatment failure, however not to a statistically significant degree (P > 0.064). Clinical outcomes following RAI can be stratified by identifying factors that are associated with initial treatment failure. These findings can help restratify patients for RAI treatment and change patient management in certain cases. Such stratification will ultimately help to optimize successful treatment outcomes and improve patient quality of life.