Gut microbiome signatures discriminate deep vein thrombosis through machine learning and metabolic analysis.

Deep vein thrombosis (DVT) remains difficult to distinguish because of its often silent presentation and the limited specificity of current diagnostic tools. We aimed to evaluate whether integrating gut microbiome profiles with routine clinical data could enhance the classification performance for identifying DVT in a case-control cohort. Stool samples were collected from individuals with DVT (n = 58), coronary artery disease (CAD, n = 56), and healthy controls (HC, n = 500). Full-length 16S rRNA gene sequencing was used to characterize the gut microbiota at species-level resolution. A random forest classifier was trained using a nested cross-validation framework, with permutation importance and SHAP (Shapley additive explanations) analyses applied to assess model interpretability. Decision curve analysis (DCA) was employed to evaluate the discriminative value of the models in an independent test set. Following linear discriminant analysis (LDA) effect size (LEfSe) screening, 95 candidate microbial features were entered into a random forest framework. Features were reduced using mutual information filtering and embedded selection to retain the final 10 for DVT vs. non-DVT classification. The integrated microbial-clinical model demonstrated substantially improved discrimination compared with the clinical-only model, achieving higher ROC-AUC [0.947 (95% CI 0.870-0.991) vs. 0.874 (95% CI 0.794-0.941)] and PR-AUC [0.793 (95% CI 0.602-0.931) vs. 0.497 (95% CI 0.274-0.724)]. Importantly, the microbiome-derived signals were robustly associated with DVT risk after adjustment for clinical covariates. Functional prediction analysis indicated enrichment of vitamin K