# Prediction of total iodine dose of I-131 therapy for Graves’ hyperthyroidism achieved remission status: a random forest regressor model approach to assess treatment efficacy

## TL;DR

This study uses a machine learning model to predict the right dose of radioactive iodine for treating Graves’ disease, improving treatment accuracy and personalization.

## Contribution

A novel Random Forest Regressor model with SHAP interpretability for predicting optimal I-131 doses in Graves’ hyperthyroidism.

## Key findings

- The RFR model achieved high accuracy with R-squared values of 0.858 and 0.838 on validation sets.
- SHAP analysis revealed key factors like FT4, Teff, and thyroid weight significantly influence predicted iodine dose.
- The model enables personalized dosing, potentially reducing adverse effects like hypothyroidism.

## Abstract

Graves’ hyperthyroidism (GH) presents significant challenges in optimizing Iodine-131 (I-131) therapy, largely due to the variability in patient responses and the limitations of traditional dosing methods. This study aimed to develop and validate a Random Forest Regressor (RFR) model to predict the effective total iodine dose (TID) necessary to achieve remission in patients with GH, thereby enhancing precision and individualization in patient management.

A retrospective cohort study design was employed, analyzing comprehensive clinical data from 975 adult GH patients who achieved remission and underwent 131I therapy 25 January 2015 and 8 August 2023. The cohort, consisting of 975 patients, was divided into a development set (n = 633, spanning from 25 January 2015 to 25 January 2021) and a temporal validation set (n = 342, covering the period from 26 January 2021 to 8 August 2023). A RFR model was developed, utilizing variables such as gender, iodine dose per gram of thyroid tissue (IDPG), Free Thyroxine (FT4), 24-hour Radioactive Iodine Uptake (RAIU24h), Effective half-life (Teff), and thyroid weight to predict the TID. The model’s interpretability was further enhanced using SHapley Additive exPlanations (SHAP) values.

Key predictive variables identified through LASSO-Gaussian regression analysis were gender, IDPG, FT4, RAIU24h, Teff, and thyroid weight. The RFR model demonstrated strong predictive performance, achieving an R-squared value of 0.858 ± 0.05 on the validation set and 0.838 on the temporal validation set, indicating its high capability to explain the variance in TID. SHAP analysis provided crucial insights into the contribution of each feature, highlighting, for example, that high FT4, Teff, and thyroid weight were primary positive contributors to the predicted TID, while RAIU24h offered a compensatory negative contribution.

In conclusion, this study successfully developed and validated an RFR model that accurately predicts the TID for GH patients achieved remission. By integrating multi-dimensional features and providing interpretability through SHAP values, this model offers a sophisticated approach to dose personalization. This advancement has the potential to significantly improve 131I treatment efficacy, minimize adverse effects such as hypothyroidism, and foster more precise, individualized patient care in GH.