Application of an Interpretable Machine Learning Model to Predict Lymph Node Metastasis in Patients with Laryngeal Carcinoma

Abstract

Objectives. A more accurate preoperative prediction of lymph node metastasis (LNM) plays a decisive role in the selection of treatment in patients with laryngeal carcinoma (LC). This study aimed to develop a machine learning (ML) prediction model for predicting LNM in patients with LC. Methods. We collected and retrospectively analysed 4887 LC patients with detailed demographical characteristics including age at diagnosis, race, sex, primary site, histology, number of tumours, T-stage, grade, and tumour size in the National Institutes of Health (NIH) Surveillance, Epidemiology, and End Results (SEER) database from 2005 to 2015. A correlation analysis of all variables was evaluated by the Pearson correlation. Independent risk factors for LC patients with LNM were identified by univariate and multivariate logistic regression analyses. Afterward, patients were randomly divided into training and test sets in a ratio of 8 to 2. On this basis, we established logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), and light gradient boosting machine (LightGBM) algorithm models based on ML. The area under the receiver operating characteristic curve (AUC) value, accuracy, precision, recall rate, F1-score, specificity, and Brier score was adopted to evaluate and compare the prediction performance of the models. Finally, the Shapley additive explanation (SHAP) method was used to interpret the association between each feature variable and target variables based on the best model. Results. Of the 4887 total LC patients, 3409 were without LNM (69.76%), and 1478 had LNM (30.24%). The result of the Pearson correlation showed that variables were weakly correlated with each other. The independent risk factors for LC patients with LNM were age at diagnosis, race, primary site, number of tumours, tumour size, grade, and T-stage. Among six models, XGBoost displayed a better performance for predicting LNM, with five performance metrics outperforming other models in the training set (AUC: 0.791 (95% CI: 0.776–0.806), accuracy: 0.739, recall rate: 0.638, F1-score: 0.663, and Brier score: 0.165), and similar results were observed in the test set. Moreover, the SHAP value of XGBoost was calculated, and the result showed that the three features, T-stage, primary site, and grade, had the greatest impact on predicting the outcomes. Conclusions. The XGBoost model performed better and can be applied to forecast the LNM of LC, offering a valuable and significant reference for clinicians in advanced decision-making.