OBJECTIVES Medical errors are a major problem in the United States. The Institute of Medicine estimates that medical errors cost the U.S. approximately $37.6 billion, of which $17 billion are preventable. Our objectives are to develop a Bayesian network (BN) model to detect value errors in blood lab and to compare performance of our model with an existing automated ruled-based approach (LabRespond), and a logistic regression model. METHODS The sample consisted of 5800 observations from the National Health and Nutrition Examination Survey dataset. The performance was assessed by the area under the receiver-operating characteristics curves (AUCs) using a 10-fold cross validation methodology. Small, medium, and large errors were randomly generated and added to liver enzymes (AST, ALT, and LDH). The outcome of interest was the correct detection of liver enzymes as “error” or “normal.” In BN, the outcome was predicted by exploiting probabilistic relationships among AST, ALT, LDH, and gender. Addition to AST, ALT, LDH, and gender, LabRespond required more analyte information (GGT, ALP, and total bilirubin) to achieve optimal prediction. For the Logistic model, the model was determined by stepwise selection among analytes that were significant at α<0.05 RESULTS The BN was predictive of added errors with small [AUC=0.644 (0.023)], medium [AUC=0.787 (0.019)], and large [AUC=0.903 (0.013)] error sizes; and performed significantly better than LabRespond [z=1.99 (p<0.05), z=2.77 (p<0.01), and z=4.57 (p<0.001), respectively] and the logistic model [z=1.71 (p<0.05), z=5.28 (p<0.001), and z=8.87 (p<0.001), respectively] CONCLUSIONS A BN model detects errors better and with less information than existing automated models, suggesting that Bayesian model can be an effective means for reducing medical costs in the laboratory