Interpretable machine learning reveals metabolomic signatures: biomarkers and mechanisms in acute vs chronic angle-closure glaucoma

Abstract Backgrounds Acute primary angle-closure glaucoma (APACG) and chronic primary angle-closure glaucoma (CPACG) exhibit distinct clinical features, yet the molecular mechanisms underlying their differing progression rates remain unclear. This study integrates metabolomics and machine learning to identify subtype-specific metabolic profiles and serum biomarkers for distinguishing APACG from CPACG. Methods A total of 128 patients were included: 47 APACG and 47 CPACG patients from the Eye & ENT Hospital of Fudan University, 20 APACG and 14 CPACG patients from Xuhui Central Hospital. Serum metabolomics was performed via UPLC-MS/MS. Differentially abundant metabolites were identified through metabolomic profiling, and machine learning models were developed to classify subtypes. Model performance was evaluated via receiver operating characteristic (ROC) curves, precision‒recall curves, and decision curve analysis. OPLS-DA revealed significant metabolic differences between APACG and CPACG, particularly in the amino acid and caffeine metabolism pathways. Results Eight differentially abundant metabolites, including caffeine and its metabolites, were consistently identified in both sets. Among the 10 machine learning models, XGBoost demonstrated the best performance, with AUC values of 0.85 (training set) and 0.82 (independent validation set). SHAP analysis highlighted 1-methylxanthine and 3-methylxanthine as key contributors to the model’s predictive performance. Conclusions This study revealed distinct metabolic profiles between APACG and CPACG, with caffeine and its metabolites playing a significant role. The XGBoost model exhibited robust predictive performance, suggesting its potential clinical utility for differentiating PACG subtypes.