Abstract
BACKGRUOUND: Osteoporosis is a common complication among thyroid cancer survivors; however, predictive tools for this condition remain inadequate. This study aimed to develop time-to-event prediction models for assessing osteoporosis risk in thyroid cancer patients. METHODS: Using the Korean National Health Insurance Service claims database, we identified 3,089 patients newly diagnosed with thyroid cancer between 2004 and 2014. Patients were randomly divided into training and test datasets in a 7:3 ratio. Three time-toevent models were constructed: random survival forest, Boruta-Cox proportional hazards, and least absolute shrinkage and selection operator (LASSO)-penalized Cox models, with feature selection and five-fold cross-validation. Model performance was evaluated using time-dependent area under the curve, Harrell's concordance index (C-index), and risk stratification analysis. RESULTS: Among thyroid cancer survivors with a median follow-up of 4.2 years, the 5-year cumulative incidence of osteoporosis was 21%. The Boruta-Cox proportional hazards model achieved the highest C-index of 0.72 (95% confidence interval [CI], 0.68 to 0.75), outperforming the random survival forest (0.68 [95% CI, 0.65 to 0.71]) and the LASSO-penalized Cox model (0.64 [95% CI, 0.61 to 0.68]). Risk stratification analysis showed that all three models significantly distinguished between low- and high-risk groups (P<0.001). CONCLUSION: This study constructed well-performing prediction models for estimating osteoporosis risk in thyroid cancer survivors, demonstrating their utility in risk stratification.