Machine-learning model of eye metastasis in nasopharyngeal carcinoma based on the AdaBoost method

Abstract

Abstract Background Nasopharyngeal carcinoma (NPC) is a common cancer of the head and neck, and the eye is a common metastatic site of NPC. This study aimed to use machine learning (ML) to establish a clinical prediction model for ocular metastasis (OM) in NPC patients. Methods We retrospectively collected clinical data from 1,855 patients with NPC who were randomized to a training set and internal test set. Patients with NPC were divided into the OM group or the non-ocular metastasis (NOM) group. Independent risk factors for NPC-related hypertension risk were screened with multivariate logistic regression models. Six ML algorithms were used, including AdaBoost (AB), logistic regression (LR), random forest (RF), multilayer perceptron (MLP), bagging (BAG), and XGBoost (XGB). The training set was used to tune the model parameters to determine the final prediction model, and the test set was used to evaluate the training model. We compared the accuracy, sensitivity, area under the ROC curve, F1 score, and specificity of the models to determine the best machine-learning algorithm for predicting the probability of OM in NPC patients. In addition, a web calculator was developed to facilitate its clinical application. Results Among these six models, the AB model had the best differential diagnostic ability (F1 score = 0.773, area under the curve = 0.995, accuracy = 0.983, sensitivity = 0.833, and specificity = 0.985). Based on this model, an online web calculator was constructed to calculate the probability of OM in NPC patients to help clinicians differentially diagnose the disease. Finally, the Shapley Supplementary Interpretation library was used to screen the five most important risk factors for OM in NPC patients: TG, Cyfra 21 1, CA199, Hb, TC, and Pathology type. Conclusion We developed a risk prediction model for OM in NPC patients using ML methods and demonstrated that the AB model performed best among six ML models. This prediction model can help to identify patients with OM from NPC, provide early and individualized diagnosis and treatment plans, protect patients from OM from NPC, and minimize the burden on society.