Application of metabolomics and MCDM approach in developing a novel strategy for disease diagnosis: A case study in Primary Sjögren's Syndrome

Primary Sjögren's Syndrome (pSS) is a complex autoimmune disease with an unclear etiology. Due to the lack of a single diagnostic gold standard, multidisciplinary and invasive examinations are often required for pSS, underscoring the urgent need for innovative non-invasive approaches to simplify the diagnostic process. Leveraging advances in machine learning and metabolomics, this study developed a novel diagnostic strategy using fecal non-targeted metabolomics data from 93 pSS patients and 42 healthy controls acquired via liquid chromatography-mass spectrometry (LC-MS). Through rigorous feature optimization with Shapley additive explanations (SHAP) and multi-criteria decision-making (MCDM), 10 pivotal differential metabolites were identified from 151 metabolites. A stacking ensemble framework integrating six machine learning models achieved exceptional diagnostic performance (AUC: 0.98, sensitivity: 0.97, specificity: 0.90, recall: 0.93, accuracy: 0.95), surpassing individual model outputs. A user-friendly visualized metabolic diagnostic system was concurrently established to enhance clinical application. These findings demonstrate that integrating machine learning into metabolomics research provides a robust, non-invasive solution for pSS diagnosis, with high generalizability and clinical applicability. This integrative approach not only addresses critical diagnostic challenges in pSS but also establishes a methodological paradigm for exploring metabolic biomarkers in other complex diseases.