Development of a machine learning-based prediction model for bipolar disorder relapse via integration of 1H-NMR metabolomics and clinical features.

BACKGROUND: Bipolar disorder (BD) is a chronic psychiatric illness characterized by recurrent episodes of depression and mania, with a global lifetime prevalence of 0.6-1 %. Patients experience high relapse rates despite treatment, while existing prediction models show limited clinical utility. METHODS: We integrated proton nuclear magnetic resonance (1H NMR) metabolomics with multidimensional clinical variables in a cohort of BD patients (n = 89). Metabolomic data were analyzed using partial least squares-discriminant analysis (PLS-DA) and orthogonal partial least squares-discriminant analysis (OPLS-DA) to identify differential metabolites. Significant metabolites were further analyzed through pathway enrichment analysis. A multimodal prediction model was then constructed using LASSO-logistic regression. RESULTS: During follow-up, 38 patients (42.7 %) experienced relapse while 51 (57.3 %) remained stable. Six significant predictors of relapse were identified: three clinical features (disease duration, negative life events, sleep disorder) and three metabolites (acetone, O-acetylglycoprotein, choline phosphate). The model demonstrated strong predictive performance (training set AUC = 0.897,95 %CI(0.831-0.942);validation set AUC = 0.846，95 %CI(0.769-0.910)).Metabolomic analysis revealed five differentially expressed metabolites enriched in phospholipid biosynthesis pathways, suggesting dysregulated membrane metabolism may underlie relapse. CONCLUSION: Our findings demonstrate the clinical value of combining metabolomic and clinical markers for BD relapse prediction, while providing new insights into disease mechanisms. This approach may facilitate personalized intervention strategies and warrants validation in larger cohorts.