Interpretable machine learning model for immune-related hyperthyroidism prediction in patients receiving immune checkpoint inhibitors: A retrospective study.

AIMS: This study aimed to develop and validate an interpretable machine learning (ML) model to predict the risk of immune-related hyperthyroidism (irHT) in patients receiving immune checkpoint inhibitors (ICIs). METHODS: A retrospective cohort study included 711 patients who received ICIs treatment at the First Affiliated Hospital of Ningbo University. Patients were randomly divided into training and validation sets in a ratio of 7:3. This study used six ML algorithms: logistic regression (LR), deep neural network (DNN), random forest classifier (RF), support vector machine (SVM), light gradient boosting machine (LGBM) and extreme gradient boosting (XGBoost) to construct irHT risk prediction models. The area under the receiver operating characteristic curve (AUC) was the main evaluation metric. Model interpretability was achieved through Shapley Additive Explanation (SHAP) and Locally Interpretable Model-Independent Explanation. RESULTS: Among these models, the LGBM model showed the highest predictive performance with an AUC of 0.788 (95% CI: 0.742–0.834) in the testing set. The Delong’s test and calibration curve indicated that the LGBM model performed better than the other models. SHAP analysis showed that free triiodothyronine, thyroid-stimulating hormone, and hypertension were the top important risk factors. CONCLUSIONS: The interpretable ML model established in this study may provide a useful reference for irHT risk assessment in patients receiving ICIs. Our findings may support individualized risk assessment in clinical practice and contribute to improving the safety of ICIs therapy.