From data to decision: an interpretable machine learning model for optimizing RAI therapy in Graves' hyperthyroidism.

Area of Science:

• Endocrinology
• Medical Informatics

Background:

• Radioactive iodine (RAI) therapy is a primary treatment for Graves' hyperthyroidism (GH).
• Current dosing strategies often result in significant treatment failure due to complex individual patient responses.

Purpose of the Study:

• To develop and validate an interpretable machine learning framework for predicting RAI therapy outcomes in GH patients.
• To identify key clinical predictors of RAI treatment success or failure.

Main Methods:

• Retrospective analysis of 1,292 GH patients treated with RAI.
• Feature selection using stepwise regression with AIC to identify nine optimal predictors.
• Comparison of six machine learning algorithms, with performance assessed by AUC, Brier score, and SHAP analysis.

Main Results:

• A 75.8% remission rate was observed in the cohort.
• Nine significant predictors were identified: gender, age, antithyroid drug history, disease duration, total iodine dose, FT4, RAIU 3h, thyroid weight, and TRAb.
• The Random Forest model achieved an AUC of 0.950 and a Brier score of 0.067, demonstrating superior predictive performance.

Conclusions:

• An interpretable machine learning framework can precisely predict RAI outcomes for Graves' hyperthyroidism.
• This tool has the potential to guide personalized dosing strategies and reduce treatment failure rates.