The value of machine learning models in predicting lung metastasis in papillary thyroid carcinoma.

OBJECTIVE: To construct a machine learning (ML) model to predict the risk of lung metastasis in papillary thyroid carcinoma (PTC). METHODS: This retrospective study included 681 PTC patients who received radioactive iodine (RAI) treatment for the first time. Among them, 71 patients were diagnosed with lung metastasis (10.4%). The patients' age, sex, stimulated thyroglobulin (s-Tg) level ahead of 131I treatment, unilateral or bilateral lobe involvement, maximum tumor diameter, number of invading tumors (extracted from postoperative pathological reports, defined as the number of structures involved by extrathyroidal extension), as well as T and N status, were evaluated. The dataset was randomly divided into a training set and a testing set in a 7:3 ratio, and least absolute shrinkage and selection operator (LASSO) regression was used for feature screening. Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM) and extreme gradient boosting (XGBoost) were constructed, and we used 10-fold cross-validation in the validation of the models, assessed through the receiver operating characteristic (ROC) curves, calibration curves, area under the precision recall curve (AUPRC) and decision curve analysis (DCA). RESULTS: The selected features in the final model included "maximum tumor diameter", "number of invading tumors", and "s-Tg". The RF model had the highest area under the curve (AUC) of 0.935, whereas the SVM had the lowest generalization ability (AUC = 0.863). In the testing of all models, LR achieved an AUC of 0.928. Although the RF model had a slightly higher AUC of 0.935, the LR model was ultimately chosen due to its superior generalization ability as indicated by its highest performance in AUPRC of 0.74 in cross-validation. DCA showed that under most threshold probabilities, the net benefit of the LR model exceeded other models. A nomogram was used to display the LR results. CONCLUSION: LASSO regression analysis identified that large primary tumor, high number of invading tumors, high s-Tg level ahead of 131I treatment might be considered as risk factors for lung metastasis in PTC patients. Among the ML models constructed accordingly, the LR model demonstrated strong performance in predicting lung metastasis of PTC.