Machine learning-based prediction of lymph node metastasis in papillary thyroid carcinoma: an interpretable multidimensional model.

BACKGROUND

Lymph node metastasis (LNM) is a prevalent feature in papillary thyroid carcinoma (PTC) that significantly influences surgical management and prognosis. Accurate identification of LNM risk is essential for optimizing adjuvant therapy and long-term surveillance strategies in clinical practice.

METHODS

This retrospective study analyzed 232 patients who underwent thyroidectomy. We employed LASSO regression to select six predictive features from a comprehensive covariate panel. Subsequently, seven machine learning algorithms were trained and evaluated via grid search and 5-fold cross-validation. The optimal model was interpreted using SHAP analysis to elucidate feature contributions.

RESULTS

LASSO-based feature selection retained six predictive features: BRAF mutation status, tumor size, capsular invasion, extrathyroidal extension, multifocality, and TSH. Among the seven algorithms, the Support Vector Machine (SVM) model showed the most favorable validation performance, achieving an AUC of 0.849 (95% CI, 0.756-0.934) and an accuracy of 0.768 in the validation cohort. SHAP analysis identified TSH and tumor size as the most influential contributors to model predictions, enabling transparent and individualized LNM risk stratification.

CONCLUSION

The SVM-based model demonstrated favorable internal validation performance and provided an interpretable framework for individualized LNM risk stratification in patients with PTC. Further prospective multicenter validation is required before routine clinical implementation.