Abstract:

Objective Construct a light gradient boosting machine （LightGBM） model based on ultrasonic and clinical features， and perform shapley additive explanation （SHAP） on the model for assessing its value in predicting cervical lymph node metastasis （CLNM） of medullary thyroid cancer （MTC）． Methods A retrospective analysis included the ultrasonic and clinical characteristics of 88 patients pathologically diagnosed with MTC. The Random Forest-based Recursive Feature Elimination （RF-RFE） algorithm with repeated cross - validation was applied to screen independent predictive variables for cervical lymph node metastasis in medullary thyroid carcinoma （MTC-CLNM）， and a LightGBM model was developed to predict MTC-CLNM. Receiver Operating Characteristic （ROC） curves， calibration curves， and decision curve analysis （DCA） were plotted to assess the predictive performance of the model. Additionally， SHAP values were introduced to conduct interpretability analysis for the predictive model， and the five-fold cross-validation method was applied to perform internal validation. Results A LightGBM model was constructed to predict MTC-CLNM， achieving a predictive accuracy of 0.795 and an AUC value of 0.852. The SHAP values analyzed the decision weights of each feature in the model， with the order from largest to smallest being： Ctn， age， and capsular invasion. The model was internally validated using the five - fold cross - validation method， and the average AUC was 0.831. Further Bootstrap validation showed that the median AUC reached 0.820， indicating good stability of the model. Conclusion The LightGBM model based on ultrasonic and clinical characteristics exhibits high diagnostic performance for predicting MTC-CLNM， and the intuitive interpretation of SHAP becomes an effective tool for guiding clinical decisions.

Key words: medullary thyroid carcinoma, cervical lymph node metastasis, LightGBM, ultrasound, SHAP values