Precision diagnosis model for treatment-resistant depression integrating serum metabolomics and clinical risk factors

Objective This study aimed to construct a high-efficiency dual-modal diagnostic model for treatment-resistant depression (TRD) by integrating serum metabolomics and clinical risk factors, and explore its metabolic pathological mechanisms. Methods A total of 93 major depressive disorder (MDD) patients (53 TRD, 40 non-TRD) were enrolled for a single-center retrospective study. Serum untargeted metabolomics and clinical baseline data were collected, with differential metabolites and clinical risk factors screened by statistical analysis and multi-step machine learning to identify core features. Five machine learning algorithms were compared to build unimodal and random forest-based dual-modal diagnostic models, and KEGG pathway enrichment analysis was performed. Results 3 core clinical risk factors (medical history, HDL, FBG) and 8 core metabolic biomarkers were identified. The dual-modal model achieved AUC 0.996 (training) and 0.911 (validation), outperforming unimodal models. Differential metabolites were mainly enriched in lipid (44.8%) and amino acid (23.9%) metabolism. Fibrinopeptide A516, 12-HETE and the three clinical factors were core driving features. Conclusion The dual-modal model has high diagnostic efficiency for TRD. TRD is associated with endocannabinoid system hypofunction and metabolic imbalance, which provides an objective diagnostic tool and new insights for TRD mechanism research and therapy development.