Metabolomic machine learning predictor for the diagnosis of alcohol-associated liver disease.

BACKGROUND: It is urgent to find novel non-invasive biomarkers that can accurately diagnose alcohol-associated liver disease (ALD). The objective of this study was to explore dysregulated metabolites in the serum of ALD patients by metabolomics, and establish a reliable diagnostic model by machine learning algorithms. METHODS: A total of 1800 participants, including ALD, metabolic dysfunction-associated steatotic liver disease (MASLD), chronic hepatitis B (CHB), alcohol use disorder (AUD) and normal control (NC) individuals were recruited from four medical centers. Steroid hormone and bile acid metabolism was identified to be dysregulated in ALD in the discovery cohort by untargeted metabolomic analysis, and further confirmed in the training cohort by absolute quantitative metabolomic analysis. A machine learning model named "Bashald" was built based on the training cohort, and further validated in three independent validation cohorts. RESULTS: Our Bashald model exhibited great diagnostic performance with an AUC of 0.942 (95% CI, 0.880-1.000) in an internal test subset. In the validation cohorts, Bashald maintained good predictive performances, with the AUCs of ≥0.821 for diagnosing ALD. In addition, Bashald demonstrated superior performance for the detection of early-stage ALD patients, with the AUCs of 0.870 and 0.792 for the training cohort and validation cohort 2, respectively, which had greatly surpassed traditional clinical indicators. CONCLUSIONS: Our research uncovered the specific metabolic profile of ALD and identified a distinct set of biomarkers that facilitate early detection, thereby promoting the application of precision diagnosis for ALD.