Interpretable Machine Learning for the Identification of Key Metabolic Biomarkers Associated with Newly Diagnosed Malignancies

Objective: To identify the most informative metabolic biomarkers associated with newly diagnosed cancer and to evaluate the potential of interpretable machine learning methods for their identification and patient classification. Materials and methods: This single-center retrospective study included 210 patients: 110 subjects without cancer and 100 patients with newly diagnosed malignancies. Clinical, anthropometric, laboratory, and metabolic variables were analyzed, including body mass index, waist circumference, visceral adiposity index, fasting glucose, immunoreactive insulin, insulin resistance indices, lipid profile, adipokines, and inflammatory markers. After preprocessing and stratified splitting into training, validation, and test sets, a family of Logistic Regression models, Elastic Net, Decision Tree, and CatBoost were used for binary classification. Model interpretation was performed using SHAP analysis, CatBoost feature importance, decision tree structure, SHAP Waterfall plots, and standardized Elastic Net coefficients. Results: CatBoost demonstrated the best classification performance on the test set (AUROC=0.9805; AUPRC=0.9711; Recall=0.9091; Precision=0.9524; F1-score=0.9302). Independent interpretation methods consistently identified the hyperglycemia criterion, fasting glucose, HOMA-IR, oral glucose tolerance test parameters, and the number of metabolic syndrome components as the leading predictors. Distribution analysis confirmed a shift of these biomarkers toward more pronounced carbohydrate metabolism disorders in the oncology group. Conclusion: Interpretable machine learning can support the identification of metabolic biomarkers associated with newly diagnosed cancer. Carbohydrate metabolism and insulin resistance markers were the most informative predictors and may be useful for early oncometabolic risk stratification and future clinical decision-support models.