Development and validation of a machine learning-based hierarchical classification model for predicting the early therapeutic response to initial 131I therapy in papillary thyroid carcinoma.

OBJECTIVE

Personalized tools for accurately predicting papillary thyroid carcinoma (PTC) patients' response to initial 131I therapy are lacking. This study aimed to develop a machine learning (ML) model for prediction of 6-12 month therapeutic response after therapy.

METHODS

This study retrospectively enrolled 702 PTC patients undergoing initial 131I therapy. A hierarchical classification framework was designed: the first layer distinguished excellent response from non-excellent response (Non-ER), while the second layer further categorized Non-ER into indeterminate response, biochemical incomplete response, and structural incomplete response. Seven ML algorithms were utilized per layer. Feature selection was performed using recursive feature elimination. Model development employed nested five-fold cross-validation within the training set, with performance further evaluated on an independent testing set. The final model was benchmarked against the American Thyroid Association risk stratification system, interpreted via Shapley Additive exPlanations, and deployed as an interactive web tool.

RESULTS

For the first-layer classification, the Random Forest model optimally identified Non-ER. For the second-layer classification, the Logistic Regression + Extreme Gradient Boosting fusion model excelled. The final model, built on 17 features, achieved a hierarchical F1 score of 0.849 and an overall accuracy of 69.7%. Both the model's superior predictive accuracy over the benchmark model and its strong association with progression-free survival confirmed the value of this hierarchical model.

CONCLUSIONS

This study developed an ML-based hierarchical model capable of forecasting early response to initial 131I therapy in PTC, offering a practical web-based tool for personalized clinical decision-making.