Development and validation of a machine learning-based model for predicting radiation-induced hypothyroidism in
Fangyan Zhong1,2,3, Wenchao Zhou1,2,3, Xiying Gao1,2,3
1NHC Key Laboratory of Personalized Diagnosis and Treatment of Nasopharyngeal Carcinoma, Jiangxi Cancer Hospital, The Second Affiliated Hospital of Nanchang Medical College, Nanchang, China.
Background And Purpose:
This study aims to develop a robust and user-friendly prediction model for radiation-induced hypothyroidism (RIHT) in nasopharyngeal carcinoma (NPC) patients.
Materials And Methods:
NPC patients treated with IMRT between Jan. 2019 and Dec. 2021 were randomly assigned to a training cohort (n = 328) and a validation cohort (n = 141) at a ratio of 7:3. A total of 33 clinical and dose-volume variables were collected. Significant variables (p < 0.05) were identified through univariate Cox analysis and further refined using a 101-combination machine learning (ML) framework to develop a robust predictive model. The model was subsequently simplified through multivariate Cox analysis and a nomogram. Finally, the performance of the model was evaluated using the C-index, calibration plots, and decision curve analysis.
Results:
Using a 101-combination ML framework, we developed a predictive model for RIHT in NPC. The Coxboost + RSF method with 11 predictors achieved the best performance (C-index: 0.91 [training], 0.71 [validation]). A simplified five-variable model (pre-treatment TSH, TSH-to-thyroid-volume ratio, age, V45, V20) was created via multi-cox regression, with a C-index of 0.80 [training] and 0.71 [validation]. High-risk patients had significantly higher three-year RIHT incidences (72.3% vs. 18.6%, p < 0.0001) in the training cohort, and 67.9% versus 24.4% (p < 0.0001) in the validation cohort. The model showed strong calibration and confirmed clinical utility through decision curve analysis, supporting its use in personalized treatment planning.
Conclusion:
We developed a ML framework to identify key predictive factors for RIHT, which was simplified into a five-variable model for clinical use, offering a robust tool for predicting RIHT risk in decision-making.
