Reservations on the Use of Error Grid Analysis for the Validation of Blood Glucose Assays

The results of the Diabetes Control and Complications Trial (1) have led to an increased emphasis on frequent blood glucose assay and on the use of blood glucose values to indicate more appropriate strategies of insulin injection. This suggests the need for the development of new and convenient blood glucose assays. This article comments on the use of "error grid analysis" (2,3) for validating novel blood glucose assays. Error grid analysis categorizes individual blood glucose values by possible clinical consequences of inaccuracy in the reported values (3). In this approach, the glucose concentration of a blood sample is determined by two methods: a new glucose assay method that gives a blood glucose estimate and a standard method taken as a blood glucose reference value that reflects the "true" blood glucose level. A point representing the sample is plotted on a graph of the type shown in Fig. 1, with the estimate on the ordinate and the reference on the abscissa (data with <50% variation around the central axis were generated for illustration). The graph is divided into five regions, each associated with a qualitatively different clinical action. In region A, the estimate lies within close proximity to the reference and appropriate action based on the estimate would not differ from that taken based on the reference. Region B is where the estimate differs substantially from the reference, but the appropriate action taken based on the estimate would be benign or no treatment would be given. In region C, the estimate and reference differ substantially and appropriate action based on the estimate would likely result in unnecessary correction or overcorrection of blood glucose. In region D, the estimate suggests normal to mild hyperglycemia that would not typically be treated, whereas the reference corresponds to hypoor hyperglycemia that requires treatment, leading to a dangerous failure to treat. In region E, the estimate and reference differ substantially and actions based on the estimate would be dangerously opposite to the treatment actually needed. Actions associated with regions A and B are clinically acceptable, but those based on regions C through E are potentially dangerous. The exact boundaries separating each region of the error grid are somewhat arbitrary (3). Thus the clinical significance of an erroneous measurement depends on where the respective values lie on the error grid. Certain investigators have claimed that error grid analysis indicates the accuracy of an assay method (3,4). In some cases, the percentage of data in each of the upper and lower regions A through E have been reported as an approach to quantifying the data. The development of error grid analysis was motivated in part in response to certain inappropriate uses of standard statistical methods such as linear regression and Pearsons product-moment correlation coefficient for evaluating error associated with individual values of blood glucose (5). An example of a more effective statistical method for this type of problem is that proposed by Bland and Altman (6). This is a simple transformation of standard statistical methods in which the difference between individual estimated and reference values is plotted against the mean of all values. This method results in a plot of