Fatma Hilal Yagin, Abdulvahap Pinar
Journal of Clinical Practice and Research - 2025;47(5):503-511
INTRODUCTION Endometrial cancer (EC) is the most common gynecologic malignancy in women worldwide, with an estimated 417,000 new cases and 97,000 deaths per year.1 The incidence of the disease is increasing, particularly in developed countries, due to rising rates of obesity, physical inactivity, and aging populations. Approximately 40-50% of EC cases result from obesity, through its association with unopposed estrogen levels and potentially insulin resistance.2 This trend is concerning, as younger patients often present with aggressive EC subtypes and poor prognosis. Ultrasound-based distinction is challenging; therefore, early and precise diagnosis is critical. The CA-125 biomarker has limited diagnostic utility, with an area under the curve (AUC) of 0.610-0.684.3 This limitation underscores the urgent need for novel diagnostic tools to improve sensitivity and specificity. Metabolomics investigates tumor metabolic patterns in EC using tools such as the National Metabolomics Data Repository (NMDR) 2024, revealing increased glucose utilization and dysregulation of amino acids and lipids. Lipidomics has identified early-warning metabolites, such as phosphatidylcholines and sphingomyelins, which may indicate endothelial failure before symptoms appear. Studies show that EC cancers consume more branched-chain amino acids and generate excess glycolytic byproducts, thereby altering pathways and enhancing resistance to apoptosis.4,5 Recent investigations have confirmed the promise of metabolomics in EC diagnosis. PELDI-MS (ProteinChip-based Surface-Enhanced Laser Desorption/Ionization Mass Spectrometry) has identified serum metabolic fingerprints with an AUC of 0.87-0.93, while other studies have reported alterations in glutamine and amino acid metabolism associated with prognosis and treatment outcomes. Machine learning, namely AutoML-XAI, has expanded these discoveries by emphasizing transparent and effective computational models.3,6,7 Our study builds directly upon these advancements by applying a suite of robust ML models and XAI to an independent cohort, aiming to both validate and refine the search for the most impactful metabolic biomarkers for EC. This biological plausibility is grounded in the well-characterized metabolic reprogramming that underpins EC pathogenesis. Metabolic disturbances in EC are driven by obesity and unopposed estrogen. Estrogen enhances lipid synthesis, while insulin resistance associated with obesity alters glucose and fatty acid metabolism. Metabolomics may reveal dysfunctions in pathways and biomarkers unique to diseases by detecting lipogenesis, hormone imbalance, and oxidative stress.8,9 Research demonstrated the ability of serum metabolic fingerprints (SMFs), tested with PELDI-MS mass spectrometry, to distinguish EC conditions from non-EC conditions.3 More recent studies show that understanding how EC patients alter sugar and fat metabolism helps doctors develop better personalized diagnoses.10 The metabolic pathways of diseases have been extensively studied through both liquid chromatography-mass spectrometry (LC-MS) and nuclear magnetic resonance (NMR) instruments, supported by automated data analysis that helps reveal disease markers. By combining powerful testing systems and information tools, researchers can better understand endometrial cancer metabolism and develop better healthcare solutions.11 The integration of powerful analytic tools with XAI methods helps scientists interpret complex data findings and better understand their biomarker results. XAI methodologies, such as SHAP (Shapley Additive Explanations), quantify the contribution of several metabolites to diagnostic predictions. This facilitates the connection between data-driven insights and clinical applicability. XAI methods boost how well we predict and understand information in SMFs to create more accurate metabolic biomarker panels. Studies show that using XAI alongside metabolomics helps doctors understand disease processes and make better treatment choices.7,12 Although previous research provides a foundation, many gaps remain. Many metabolomic models function as "black boxes," restricting biological understanding, while approaches such as LightGBM and XGBoost are underexplored with EC data. The transition from m/z patterns to biological significance is often neglected. To overcome this, we incorporate XAI from the start, using SHAP to quantify metabolite contributions and ensure interpretability. We compare Random Forest, BaggedCART, AdaBoost, XGBoost, and LightGBM to determine the most robust model, with the goal of developing not only a dependable diagnostic tool but also an interpretable framework that emphasizes critical metabolic indicators and promotes clinical validation.
