Exploring risk factors for cervical lymph node metastasis in papillary thyroid microcarcinoma: construction of a novel population-based predictive model

Abstract

Background: Machine learning was a highly effective tool in model construction. We aim to establish a machine learning-based predictive model for predicting the cervical lymph node metastasis (LNM) in papillary thyroid microcarcinoma (PTMC).

Methods: We obtained data on PTMC from the SEER database, including 10 demographic and clinicopathological characteristics. Univariate and multivariate logistic regression (LR) analyses were applied to screen the risk factors for cervical LNM in PTMC. Risk factors with P < 0.05 in multivariate LR analysis were used as modeling variables. Five different machine learning (ML) algorithms including extreme gradient boosting (XGBoost), random forest (RF), adaptive boosting (AdaBoost), gaussian naive bayes (GNB) and multi-layer perceptron (MLP) and traditional regression analysis were used to construct the prediction model. Finally, the area under the receiver operating characteristic (AUROC) curve was used to compare the model performance.

Results: Through univariate and multivariate LR analysis, we screened out 9 independent risk factors most closely associated with cervical LNM in PTMC, including age, sex, race, marital status, region, histology, tumor size, and extrathyroidal extension (ETE) and multifocality. We used these risk factors to build an ML prediction model, in which the AUROC value of the XGBoost algorithm was higher than the other 4 ML algorithms and was the best ML model. We optimized the XGBoost algorithm through 10-fold cross-validation, and its best performance on the training set (AUROC: 0.809, 95%CI 0.800-0.818) was better than traditional LR analysis (AUROC: 0.780, 95%CI 0.772-0.787).

Conclusions: ML algorithms have good predictive performance, especially the XGBoost algorithm. With the continuous development of artificial intelligence, ML algorithms have broad prospects in clinical prognosis prediction.

Keywords: Conventional regression model; Machine learning; Papillary thyroid microcarcinoma cervical lymph node metastasis; Prediction model; Risk factors.

Fig. 1

Flow chart of patients selection and study design

Fig. 2

Pearson correlation test for variables

Fig. 3

ROC curves, forest plot and nomogram of the LR model for cervical LNM in PTMC. Note: A shows the ROC curves of the multivariate LR. B shows the forest plot of the multivariate LR model. C shows the risk nomogram of the multivariate LR model

Fig. 4

Model performance evaluation of different ML methods. Note: A showed the ROC curve of 5 different ML models in training set; B showed the ROC curve of 5 different ML models in validation set; C showed the AUC score forest plot of each model; D showed the reliability curve of each model

Fig. 5

Optimization and visualization of the XGBoost model. Note: A, B and D displayed the ROC curve of the train, validation and test of the XGBoost model by 10-fold cross-validation. C showed the learning curve of the XGBoost classifier. E showed the reliability curve of XGBoost model. F showed the summary plots of SHAP values for the XGBoost model. For each feature, one point corresponds to a single patient. A point’s position along the x axis represented the impact that feature had on the model’s output for that specific patient. Features were arranged along the y axis based on their importance, which was given by the mean of their absolute Shapley values. The higher the feature was positioned in the plot, the more important it was for the model