A machine learning model for predicting bone and/or lung metastasis in differentiated thyroid carcinoma: enhancing precision in risk stratification.

Background: Differentiated thyroid cancer (DTC) incidence is rapidly rising worldwide. While most cases have a favorable prognosis, a subset of patients develop aggressive disease with distant metastases, particularly to the bone and lung, which significantly worsens outcomes. Current prediction models are limited in accuracy, often relying on basic clinical factors. This study aims to develop a machine learning model to improve prediction of bone and lung metastasis in DTC, enhancing risk stratification and early intervention. Methods: Using the SEER database, we developed several machine learning models-including XGBoost, Random Forest, Gradient Boosting Machine, Logistic Regression, Naive Bayes, and Classification and Regression Trees (CART)-to predict bone and lung metastasis risk in DTC patients. LASSO regression was applied to select key predictive variables, and SMOTE was used to address data imbalance. The model's generalizability was evaluated using an external validation cohort from China. Results: The XGBoost model demonstrated the highest performance, achieving an AUC of 0.988. Key predictive variables identified and included in the model were tumor size, radiation therapy, surgical interventions, histologic types, T and N stages, laterality, race, and household income. SHAP analysis confirmed the importance of these variables, with tumor size, radiation, and surgery emerging as primary predictors. In the external validation cohort, the model achieved an AUC of 0.866, indicating reliable predictive capability across clinical settings. Conclusion: This model accurately predicts bone and lung metastasis risk in DTC, offering valuable clinical utility for risk stratification and supporting early intervention strategies to improve outcomes in high-risk patients.