Continuous glucose monitoring in newborn infants: how do errors in calibration measurements affect detected hypoglycemia?

Abstract

Neonatal hypoglycemia is common and can cause serious brain injury. Continuous glucose monitoring (CGM) could improve hypoglycemia detection, while reducing blood glucose (BG) measurements. Calibration algorithms use BG measurements to convert sensor signals into CGM data. Thus, inaccuracies in calibration BG measurements directly affect CGM values and any metrics calculated from them. The aim was to quantify the effect of timing delays and calibration BG measurement errors on hypoglycemia metrics in newborn infants. Data from 155 babies were used. Two timing and 3 BG meter error models (Abbott Optium Xceed, Roche Accu-Chek Inform II, Nova Statstrip) were created using empirical data. Monte-Carlo methods were employed, and each simulation was run 1000 times. Each set of patient data in each simulation had randomly selected timing and/or measurement error added to BG measurements before CGM data were calibrated. The number of hypoglycemic events, duration of hypoglycemia, and hypoglycemic index were then calculated using the CGM data and compared to baseline values. Timing error alone had little effect on hypoglycemia metrics, but measurement error caused substantial variation. Abbott results underreported the number of hypoglycemic events by up to 8 and Roche overreported by up to 4 where the original number reported was 2. Nova results were closest to baseline. Similar trends were observed in the other hypoglycemia metrics. Errors in blood glucose concentration measurements used for calibration of CGM devices can have a clinically important impact on detection of hypoglycemia. If CGM devices are going to be used for assessing hypoglycemia it is important to understand of the impact of these errors on CGM data.

Keywords: algorithm; continuous glucose monitor; hypoglycemia; intensive care unit; neonatal; recalibration.

Figure 1.

Distribution of time delay data from both Waikato (left) and Christchurch (right) data sets. Waikato Hospital data had 1947 time delay values and an exponential decay model (µ = 12.96) fit the empirical data. Christchurch Hospital had 155 delay values and an exponential decay model (µ = 8.84) captured the empirical data.

Figure 2.

Example continuous glucose monitoring (CGM) traces showing the effect of (A) Abbott measurement error, (B) Nova measurement error, and (C) Roche measurement error. The colored band in each plot shows the 5th-95th percentile variation in the CGM trace over 1000 MC simulations.

Figure 3.

Duration of hypoglycemia in each of 1000 MC runs plotted against baseline duration of hypoglycemia, for all 155 CGM data sets. (A) Simulation results using Abbott measurement error. (B) Simulation results using Roche measurement error.