Detecting Errors in Blood Labs Using Bayesian Networks

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): The Institute of Medicine seminal report on medical errors highlighted the urgency of their identification. Medical errors are estimated to cost the U.S. between $17 billion and $29 billion a year. Clinical laboratories provide about 70% of the data used to make clinical decisions and produce an estimated 70 million errors per year in the U.S. Current methods for detecting clinical laboratory errors can be improved. Our hypothesis is that a Bayesian approach will improve error detection. The long-term objective of this project is to evaluate if Bayesian networks are more accurate than laboratory experts in detecting errors in clinical laboratory data. We use non-diabetic, pre-diabetic and diabetic clinical trial blood panel data as models for this research. The specific aims of this proposal are: (1) To construct and validate a Bayesian belief network designed to detect errors in the clinical laboratory values. One that expands on our preliminary work to include other factors that influence measured values. To accomplish this aim we will extract and clean a data set from a randomized controlled trial investigating diabetes treatments. We randomly split the data into a training set and a test set, insert errors into each data set in ways analogous to how they would be rendered naturally and validate a Bayesian belief network from the training data using a 10-fold cross validation. By varying the probability threshold used to classify data as erroneous, we will determine the sensitivity and specificity of the network as well as the area under the receiver operating characteristics curve. Finally, network vs. human expert performance will be compared on measures of sensitivity, specificity, and (z-critical) statistical differences between areas under the receiver operating characteristics curves. (2) To determine whether the Bayesian network in Aim 1 generalizes to pre- and non-diabetic populations. We will test whether the network structure in Aim 1, is effective in detecting laboratory errors in more general data sets with only parameter learning. We will perform a 10-fold cross-validation over learned network parameter estimates in each of a pre- and a non-diabetic data set. We will determine the sensitivity and specificity of the network in each data set as well as the (z-critical) statistical differences between areas under the receiver operating characteristics curve. Finally, network vs. human expert performance will be compared on the aforementioned measures. The project's health-relatedness is evident by its goal of reducing clinical laboratory errors that can adversely affect the health of healthcare recipients. The success of this project will result in the development of a method that clinical laboratories may use to detect errors in practice and save both lives and substantial health-care resources. By reducing errors in the clinical laboratory, lives and substantial health-care resources will be saved.