Development of a Machine Learning-based Model for Methimazole Dosage Adjustment in Youth With Hyperthyroidism

CONTEXT: Accurate methimazole (MMI) dose adjustment in pediatric hyperthyroidism is crucial, but individualized titration relies on clinician experience due to a lack of validated predictive tools. OBJECTIVE: This study aimed to develop and validate machine learning-based models for predicting optimal MMI dosage in pediatric hyperthyroidism. DESIGN: This was a retrospective, multicenter, model-development study. Machine learning models, including linear regression, decision tree, support vector regression, Extreme Gradient Boosting (XGBoost), and feed-forward neural networks, were trained and validated. SETTING: Data were collected from a primary center for model training, with 2 separate centers providing data for external validation. PATIENTS OR OTHER PARTICIPANTS: Data were derived from 1512 visits for the training set and 666 and 31 visits for 2 external validation cohorts, respectively. All data were from youth aged ≤18 years with hyperthyroidism. INTERVENTIONS: The models were trained to predict the optimal daily dosage of MMI based on variables including age, sex, anthropometric measures, prior MMI dosage, treatment duration, and current and previous results of thyroid function tests. MAIN OUTCOME MEASURES: Model performance was evaluated by the mean absolute error (MAE) between the predicted and actual MMI dosages. Feature importance was determined using Shapley additive explanations (SHAP) analysis. RESULTS: The XGBoost model demonstrated the best performance in both internal validation (MAE 1.72 mg) and external validation (MAE 1.08 mg). SHAP analysis identified previous MMI dose, T3, and free T4 levels as key predictors. CONCLUSION: This study introduces the first data-driven tool to guide MMI dosing in pediatric hyperthyroidism, which can improve clinical efficiency.