Abstract
<title>Abstract</title> <p>Background We aimed to develop and validate a machine learning (ML) model based on routine multidimensional laboratory indicators for individualized prediction of distant metastasis status in patients with newly diagnosed nasopharyngeal carcinoma (NPC). Methods A total of 1,085 patients with newly diagnosed NPC were retrospectively enrolled. Systematic integration of five categories of laboratory parameters—complete blood count, biochemistry, coagulation, immunology, and tumor markers—was performed. Three algorithms, including XGBoost, Random Forest, and ElasticNet, were employed for model construction and comparison. Model robustness was assessed through stratified 5-fold cross-validation and 1,000 Monte Carlo cross-validations (MCCVs). SHapley Additive exPlanations (SHAP) analysis was used to elucidate feature importance. Multilevel independent validation was conducted using multivariable logistic regression, tertile-based risk stratification, external validation with the Surveillance, Epidemiology, and End Results (SEER) database (N ≈ 7,580), and Gene Expression Omnibus (GEO) transcriptomic validation. Results The XGBoost model incorporating the comprehensive feature set achieved optimal predictive performance, with an area under the receiver operating characteristic curve (AUROC) of 0.8595 ± 0.0231 and a Brier score of 0.1287. The 1,000 MCCV repetitions confirmed an AUROC of 0.8470 (95% confidence interval [CI]: 0.8032–0.8871), excluding overfitting. SHAP analysis identified red cell distribution width-standard deviation (RDW-SD; 0.452), immunoglobulin A (IgA; 0.347), carbohydrate antigen 125 (CA125; 0.301), and D-dimer (0.284) as the core predictive factors. Risk stratification demonstrated a significant dose-response relationship: low-risk 7.3% → intermediate-risk 22.2% → high-risk 56.5% (Cochran-Armitage trend test Z = 11.42, P < 0.0001). Multivariable logistic regression confirmed aspartate aminotransferase (AST; odds ratio [OR] = 2.70), red cell distribution width (RDW; OR = 1.34), and hemoglobin (Hb; OR = 0.77) as independent predictors. SEER validation confirmed the strong prognostic effect of distant metastasis (overall survival [OS] hazard ratio [HR] = 3.21, with 23 of 24 subgroups showing significant HRs), and the nomogram achieved a C-index of 0.752. GEO transcriptomic validation revealed that Hallmark pathways corresponding to the core predictive factors showed a systematic trend of enrichment at the nominal P-value level in GSE103611, including glycolysis (normalized enrichment score [NES] = + 1.42), E2F targets (NES = + 1.27), and interferon-gamma response (NES = + 1.32). In GSE299775, single-sample gene set enrichment analysis (ssGSEA) validation identified a nominal P-value level difference in the immunoglobulin synthesis pathway (P = 0.042). Conclusions The ML model constructed in this study effectively predicts distant metastasis in NPC based on routine laboratory tests, provides multilevel biological mechanism interpretation for the core predictive factors, and has the potential to serve as an individualized risk assessment tool to complement tumor-node-metastasis (TNM) staging, offering evidence-based support for intensified screening of high-risk patients and optimization of treatment decision-making.</p>