SUN-204 Development and validation of a machine learning-based model for methimazole dosage adjustment in children and adolescents with hyperthyroidism

Abstract Disclosure: J. Kim: None. Y. Choi: None. S. Kim: None. J. Suh: None. E. Hwang: None. E. Lee: None. H. Chae: None. Background: Accurate methimazole dosage adjustment is essential in pediatric hyperthyroidism to prevent adverse effects and achieve optimal outcomes. Traditional methods, largely reliant on clinical experience, often lack objectivity. This study aimed to develop and validate a machine learning-based model to predict methimazole dosage in children and adolescents. Methods: This retrospective study included 120 children and adolescents (1,512 records) treated with methimazole for hyperthyroidism at a pediatric clinic between July 2005 and February 2024 as the training set. The test set consisted of 55 children and adolescents (666 records) treated at another clinic from January 2007 to February 2024. Input variables included age, sex, height standard deviation score (SDS), weight SDS, body mass index (BMI) SDS, T3, free T4, Thyroid stimulating hormone (TSH), follow-up interval, and previous methimazole dose. Prediction models were developed using linear regression, Extreme gradient boosting (XGBoost), support vector regression (SVR), and feed-forward network (FFN) algorithms. Internal validation used 5-fold cross-validation, while external validation was conducted with the independent test set. SHapley Additive exPlanations (SHAP) analysis evaluated the contribution of input variables. Results: The mean methimazole dose was 8.21 ± 7.02 mg in the training set and 8.46 ± 7.7 mg in the test set. Among the models, XGBoost achieved the lowest mean absolute error (MAE) (1.97 mg in internal validation; 2.05 mg in external validation). Linear Regression demonstrated MAEs of 2.09 mg (internal validation) and 2.05 mg (external validation). SVR and FFN had slightly higher MAEs, with SVR at 2.29 mg (internal validation) and 2.36 mg (external validation), and FFN at 2.13 mg (internal validation) and 2.30 mg (external validation). SHAP analysis revealed that the previous dose, T3, free T4, and TSH were the most influential variables, followed by follow up interval, age, and BMI SDS.Conclusion: We developed a machine learning model for methimazole dosage adjustment in pediatric hyperthyroidism and validated it with an independent dataset. The findings highlight the potential of machine learning to enhance precision of treatment dosing in children and adolescents with hyperthyroidism. Keywords: Hyperthyroidism, Methimazole, Machine Learning. Presentation: Sunday, July 13, 2025