Bile acid diarrhoea (BAD) is debilitating yet treatable, but it remains underdiagnosed due to challenging diagnostics. We developed a blood test-based method to guide BAD diagnosis. We included serum from 50 treatment-naive patients with BAD diagnosed by gold standard 75selenium homotaurocholic acid test, 56 feature-matched controls and 37 patients with non-alcoholic fatty liver disease (NAFLD). Metabolomes were generated using mass spectrometry covering 1295 metabolites and compared between groups. Machine learning was used to develop a BAD Diagnostic Score (BDS). Metabolomes of patients with BAD significantly differed from controls and NAFLD. We detected 70 metabolites with a discriminatory performance in the discovery set with an area under receiver-operating curve metric above 0.80. Logistic regression modelling using concentrations of decanoylcarnitine, cholesterol ester (22:5), eicosatrienoic acid, L-alpha-lysophosphatidylinositol (18:0) and phosphatidylethanolamine (O-16:0/18:1) distinguished BAD from controls with a sensitivity of 0.78 (95% CI 0.64 to 0.89) and a specificity of 0.93 (95% CI 0.83 to 0.98). The model was independent of covariates (age, sex, body mass index) and distinguished BAD from NAFLD irrespective of fibrosis stage. BDS outperformed other blood test-based tests (7-alpha-hydroxy-4-cholesten-3-one and fibroblast growth factor 19) currently under development. BDS derived from serum metabolites in a single-blood sample showed robust identification of patients with BAD with superior specificity and sensitivity compared with current blood test-based diagnostics.