Objective To investigate the value of Shapley additive explanations (SHAP) in extreme gradient boosting (XGBoost) ultrasound model in predicting papillary thyroid carcinoma (PTC) larger than 1.0 centimeter. Methods The ultrasound data of 138 cases with 145 pieces of pathologically confirmed PTC were analyzed retrospectively and compared with 127 cases with 141 pieces of nodular goiter (NG). The ultrasound risk signs were performed scoring and analyzed by chi-square test. The data were randomly split into a training set and a testing set in the ratio of 8 : 2. XGBoost was used to construct the machine learning model on the training set, and the area under the ROC curve (AUC) of receiver operating characteristic (ROC) was used in the testing set to evaluate the efficacy of the model to predict PTC. Model interpretation was performed through SHAP values to clarify the weight of each ultrasound sign in the diagnosis of PTC. Results Among the 145 PTC and 141 NG, the AUC value of the XGBoost models were 0.941and 0.921 in the training and testing set, respectively, which were based on these 5 signs: fuzzy/irregular margins/extra-glandular invasion、UGSR＜0.83、solidity、aspect ratio (A/T) > 1 and microcalcification. The analysis of SHAP values on XGBoost model showed that the mean value of absolute SHAP values of ultrasound signs was from 0.3 to 1.3, and the weights from large to small were UGSR＜0.83、solidity、fuzzy/irregular margins/extra-glandular invasion、microcalcification, and A/T > 1. All of them were positive contribution. Conclusion The analysis of the constructed XGBoost prediction model using SHAP values can quantify and visualize the diagnostic efficacy of each sign and provide a certain basis for the improvement of diagnostic criteria for thyroid nodules.