Skip to content
Open access

Determination of specific salivary biochemical profiles in oral and ovarian cancer using machine learning and SHAP analysis

Oct 2026 · Eurasian Journal of Medicine and Oncology · 0 citations

Abstract

Introduction: Saliva contains local and systemic molecular signals that may support non-invasive cancer detection, but the diagnostic value of integrated salivary profiles across anatomically distinct malignancies remains unclear. Objective: This study evaluates whether multivariable salivary signatures could distinguish controls, ovarian cancer, and oral cavity cancer. Methods: Saliva samples from 75 controls, 89 patients with ovarian cancer, and 49 patients with oral cavity cancer were analyzed for amino acids and nitrogen metabolites, inorganic ions, cytokines, growth factors, tumor-associated markers, lipid peroxidation products, and Fourier-transform infrared (FTIR) features. Results from regularized multinomial logistic regression, random forest, and XGBoost models were compared using repeated stratified five-fold cross-validation within the training set. Final model performance was then evaluated on a held-out test set using accuracy, Cohen’s κ, balanced accuracy, macro-F1, multiclass receiver operating characteristic–area under the curve, and log loss. Permutation testing, bootstrap uncertainty estimation, SHapley Additive exPlanations (SHAP) analysis, and decision curve analysis were additionally performed. Results: XGBoost showed the strongest performance and correctly classified all ovarian and oral cavity cancer cases in the held-out test set. Permutation testing confirmed that performance exceeded that of randomly assigned class labels. SHAP analysis indicated that mainly cytokine and tumor-associated markers drove ovarian cancer predictions, whereas lipid-related FTIR features dominated oral cavity cancer predictions. Decision curve analysis demonstrated positive net benefit across threshold ranges. Conclusion: Integrated salivary profiling enabled accurate internal differentiation of local and distant malignancies. The findings support biologically distinct salivary signatures, but external validation is required.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.