Building a prediction model for nasopharyngeal carcinoma based on machine learning using laboratory test indicators: a study targeting populations in high-risk areas

Early screening of nasopharyngeal carcinoma (NPC) is of vital importance for improving the survival rate of patients with NPC. In regions with a high incidence of NPC, Epstein-Barr Virus (EBV) testing is the main screening method, but it has a relatively high false-positive rate and high economic cost. This study aimed to use machine-learning methods to construct an NPC risk prediction model based on routine laboratory test results, thereby facilitating personalized screening for populations in medium- and high-risk areas. Data from 9957 participants at two centers were analyzed: Wuzhou Red Cross Hospital (3660 patients with NPC and 3576 controls) and The People's Hospital of Guangxi Zhuang Autonomous Region (1446 patients with NPC and 1275 controls). Twenty-two candidate variables were examined, including 19 routine laboratory measurements and three derived ratios (NEUT/WBC, NEUT/LYMPH, and ALB/ALP). The Wuzhou Red Cross Hospital cohort was partitioned into training and internal test sets, whereas the People's Hospital of Guangxi Zhuang Autonomous Region cohort served as an independent external-validation set. LASSO regression was applied for predictor selection. Eight machine-learning algorithms were subsequently developed: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Naive Bayes (NB), K-Nearest Neighbor (KNN), and Multi-Layer Perceptron (MLP). The area under the curve (AUC), Accuracy (ACC), Recall, and F1 scores were used to evaluate model performance. Shapley values were used to identify important contributing variables. After screening, six core predictive variables, including gender, age, HGB, LYMPH, Ca, and Mg, were obtained. Among all models, the random forest model showed strong overall performance, with an AUC of 0.99 (95% CI 0.98–0.99) in the training set and 0.93 (95% CI 0.92–0.94) in the test cohort. In the external set, robust classification performance was also maintained (AUC = 0.83). In addition, SHAP analysis showed that LYMPH contributed most to model prediction, followed by gender, age, Ca, Mg, and HGB. Constructing a prediction model based on routine physical-examination indicators using machine learning may help identify individuals at increased risk of nasopharyngeal carcinoma and support early detection and targeted intervention in medium- and high-risk areas.