Characterizing accuracy and precision of glucose sensors and meters

Abstract

There is need for a method to describe precision and accuracy of glucose measurement as a smooth continuous function of glucose level rather than as a step function for a few discrete ranges of glucose. We propose and illustrate a method to generate a "Glucose Precision Profile" showing absolute relative deviation (ARD) and /or %CV versus glucose level to better characterize measurement errors at any glucose level. We examine the relationship between glucose measured by test and comparator methods using linear regression. We examine bias by plotting deviation = (test - comparator method) versus glucose level. We compute the deviation, absolute deviation (AD), ARD, and standard deviation (SD) for each data pair. We utilize curve smoothing procedures to minimize the effects of random sampling variability to facilitate identification and display of the underlying relationships between ARD or %CV and glucose level. AD, ARD, SD, and %CV display smooth continuous relationships versus glucose level. Estimates of MARD and %CV are subject to relatively large errors in the hypoglycemic range due in part to a markedly nonlinear relationship with glucose level and in part to the limited number of observations in the hypoglycemic range. The curvilinear relationships of ARD and %CV versus glucose level are helpful when characterizing and comparing the precision and accuracy of glucose sensors and meters.

Keywords: Bland–Altman plot; Clarke error grid; Parkes error grid; accuracy; bias; continuous glucose monitoring; diabetes mellitus type 1; diabetes mellitus type 2; glucose monitoring; mean absolute relative deviation (MARD); precision; precision absolute relative deviation (PARD); self-monitoring of blood glucose; statistical analysis.

Figure 1.

Relationship of observed glucose to comparator showing the least-squares regression line (black) and line of identity (blue).

Figure 2.

Relationship of deviations versus comparator glucose. The arithmetic (signed) deviations can vary in magnitude (bias) and in terms of their own variability depending on glucose level.

Figure 3.

The absolute deviation (AD) of the test method from the comparator shows large random sampling variability. The magnitude of the AD and its own variability depend on glucose level. The least squares regression line (black) and a third order polynomial (pink) are shown. In this example, the expected AD reaches a nadir when glucose is ~100 mg/dL and the linear relationship does not provide a satisfactory curve fit below that level.

Figure 4.

Expected absolute relative deviation (ARD) as a function of glucose level calculated from data displayed in Figures 1 -3 (black), after curve smoothing using running average of ARD combined with a fourth order polynomial curve fit. Mean ARD (MARD) is also shown as a step function for 3 ranges of glucose (40-70, 70-180, and 180-500 mg/dL) (pink), and for 4 ranges of glucose when the target range is divided into 2 segments (70-125 and 125-180 mg/dL) (blue).