A multi-variable machine learning model integrating stage, histology, grade, and treatment to predict mortality in salivary gland cancer: A SEER 2010-2021 population-based study

Background: Salivary gland cancers are rare, heterogeneous tumors with widely variable clinical outcomes. Current prognostic tools rely on limited clinicopathologic features and do not incorporate complex interactions between tumor biology and treatment patterns. We aimed to develop and validate a machine learning (ML) model integrating stage, histology, grade, tumor size, and treatment modalities to predict all-cause mortality using a large U.S. population-based cohort. Methods: We identified patients with primary malignant salivary gland tumors in the SEER database (2010-2021). Variables included age, sex, race, AJCC stage, histologic subtype, tumor size, grade, nodal status, surgery, radiation, and chemotherapy. Missing data were imputed using iterative multivariate imputation. Models evaluated included logistic regression, random forest, gradient boosting, and XGBoost. Performance was assessed with 70/30 train-test split and temporal validation using diagnoses from 2010-2017 as training and 2018-2021 as testing. Model interpretability was examined using SHAP values. Results: A total of 12,678 salivary gland cancer cases met inclusion criteria. XGBoost showed higher performance, with an AUC of 0.87 in the test set and 0.85 in temporal validation, outperforming logistic regression (AUC 0.76). Sensitivity, specificity, and PPV for high-risk classification were 0.81, 0.78, and 0.72, respectively. SHAP analysis identified stage, tumor size, histologic subtype, grade, and receipt of surgery as the most influential predictors of mortality. The model revealed nonlinear interactions between tumor grade and nodal status that were not captured by traditional regression. Conclusion: We came up with a robust, interpretable machine learning model integrating key clinical, pathologic, and treatment variables to predict mortality in salivary gland cancer with high accuracy. This tool provides a strong framework for individualized risk stratification and may help decision-making regarding treatment intensity, surveillance, and survivorship care. Further prospective validation is warranted to support clinical adoption.