Detecting 'wrong blood in tube' errors: Evaluation of a Bayesian network approach.

Abstract

In an effort to address the problem of laboratory errors, we develop and evaluate a method to detect mismatched specimens from nationally collected blood laboratory data in two experiments. In Experiments 1 and 2 using blood labs from National Health and Nutrition Examination Survey (NHANES) and values derived from the Diabetes Prevention Program (DPP) respectively, a proportion of glucose and HbA1c specimens were randomly mismatched. A Bayesian network that encoded probabilistic relationships among analytes was used to predict mismatches. In Experiment 1 the performance of the network was compared against existing error detection software. In Experiment 2 the network was compared against 11 human experts recruited from the American Academy of Clinical Chemists. Results were compared via area under the receiver-operator characteristic curves (AUCs) and with agreement statistics. In Experiment 1 the network was most predictive of mismatches that produced clinically significant discrepancies between true and mismatched scores ((AUC of 0.87 (±0.04) for HbA1c and 0.83 (±0.02) for glucose), performed well in identifying errors among those self-reporting diabetes (N=329) (AUC=0.79 (±0.02)) and performed significantly better than the established approach it was tested against (in all cases p<.0.05). In Experiment 2 it performed better (and in no case worse) than 7 of the 11 human experts. Average percent agreement was 0.79 and Kappa (κ) was 0.59, between experts and the Bayesian network. Bayesian network can accurately identify mismatched specimens. The algorithm is best at identifying mismatches that result in a clinically significant magnitude of error.