Abstract
BACKGROUND: Owing to the limited characterization of lymph nodes around the entrance point of the recurrent laryngeal nerve (LN-epRLN) in clinical lymph node negative (cN0) papillary thyroid carcinoma (PTC), this study sought to develop machine learning (ML) models to predict LN-epRLN metastasis, identify the optimal model, and improve interpretability using explainable artificial intelligence techniques. METHODS: We retrospectively reviewed 1,800 patients with cN0-PTC who underwent central lymph node dissection (CLND) with systematic LN-epRLN sampling. Histopathological evaluation confirmed metastatic status. Patients were randomly divided into training and testing sets at a 7:3 ratio. Nine ML models were constructed and optimized through 10-fold cross-validation and grid search. Performance was assessed using 11 metrics, including AUC, accuracy, sensitivity, and specificity. The best-performing model was compared against traditional nomograms via probability-based ranking analysis (PMRA). RESULTS: LN-epRLNs were identified in 149 out of 1800 PTC patients, with a metastasis rate of 19.46%. The Random Forest (RF) model outperformed others, achieving training/testing scores of 0.914/0.911 accuracy, 0.956/0.919 AUC, 0.993/0.974 specificity, and 0.609/0.500 sensitivity. A simplified model incorporating seven key predictors-total central lymph node metastasis number and ratio, pretracheal lymph node metastasis number and ratio, tumor size, age, and paratracheal lymph node metastasis number-retained high predictive performance. SHAPley Additive exPlanations (SHAP) analysis highlighted central compartment metastasis burden (number and ratio) as the most influential predictors. CONCLUSION: The interpretable ML model developed in this study, leveraging the RF, provides a reliable tool for preoperative and intraoperative prediction of LN-epRLN metastasis in cN0 PTC patients. This approach has the potential to guide individualized surgical planning, optimizing the balance between oncological resection completeness and functional preservation.