Early postoperative prediction of the risk of distant metastases in medullary thyroid cancer

Abstract

PurposeThe purpose of this study was to develop and validate a nomogram for estimating the risk of distant metastases (DM) in the early postoperative phase of medullary thyroid cancer (MTC).Patients and methodsWe retrospectively reviewed cases of patients diagnosed with MTC from the Surveillance, Epidemiology, and End Results (SEER) database from 2007 to 2017. In addition, we gathered data on patients who diagnosed as MTC at Department of Thyroid Surgery in the First Hospital of Jilin University between 2009 and 2021. Four machine learning algorithms were used for modeling, including random forest classifier (RFC), gradient boosting decision tree (GBDT), logistic regression (LR), and support vector machine (SVM). The optimal model was selected based on accuracy, recall, specificity, receiver operating characteristic curve (ROC), and area under curve (AUC). After that, the Hosmer-Lemeshow goodness-of-fit test, the brier score (BS) and calibration curve were used for validation of the best model, which allowed us to measure the discrepancy between the projected value and the actual value.ResultsThrough feature selection, we finally clarified that the following four features are associated with distant metastases of MTC, which are age, surgery, primary tumor (T) and nodes (N). The AUC values of the four models in the internal test set were as follows: random forest: 0.8786 (95% CI, 0.8070-0.9503), GBDT: 0.8402 (95% CI, 0.7606-0.9199), logistic regression: 0.8670(95%CI,0.7927-0.9413), and SVM: 0.8673 (95% CI, 0.7931-0.9415). As can be shown, there was no statistically significant difference in their AUC values. The highest AUC value of the four models were chosen as the best model since. The model was evaluated on the internal test set, and the accuracy was 0.84, recall was 0.76, and specificity was 0.87. The ROC curve was drawn, and the AUC was 0.8786 (95% CI, 0.8070-0.9503), which was higher than the other three models. The model was visualized using the nomogram and its net benefit was shown in both the Decision Curve Analysis (DCA) and Clinical Impact Curve (CIC).ConclusionProposed model had good discrimination ability and could preliminarily screen high-risk patients for DM in the early postoperative period.