Radiomics-Based machine learning for preoperative prediction of Large-Number central lymph node metastasis in papillary thyroid carcinoma.

PURPOSE: This study aimed to develop interpretable machine learning (ML) models integrating preoperative clinical, ultrasound (US), and radiomic features for predicting large-number central lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC).

METHODS

This retrospective study analyzed 1,400 PTC patients undergoing thyroidectomy with central lymph node dissection (CLND) between January 2018 and May 2025. Stratified by large-number CLNM status, patients were randomized into training (n = 980) and test (n = 420) cohorts. Key predictors were identified via Least Absolute Shrinkage and Selection Operator (LASSO) and logistic regression. Six ML algorithms were trained using Synthetic Minority Over-sampling Technique (SMOTE) for class imbalance. Performance evaluation used the area under the receiver operating characteristic curve (ROC-AUC), precision-recall curve AUC (PR-AUC), and decision curve analysis (DCA). Optimal model interpretation used Shapley Additive Explanations (SHAP).

RESULTS

Large-number CLNM occurred in 92 (6.6%) patients. Predictors included clinical/US features (age, tumor size, suspicious nodes, tracheal capsular invasion) and a 7-feature LASSO-selected radiomics signature. The integrated model achieved ROC-AUCs of 0.880 (95% confidence interval: 0.841–0.912; training) and 0.827 (0.738–0.901; testing), significantly outperforming submodels in PR-AUC (p < 0.05). SHAP quantified feature contributions.

CONCLUSION

This ML model preoperatively identifies PTC patients at high risk for large-number CLNM, offering potential to personalize treatment decisions and reduce unnecessary surgery.