Page 190 - Glucose Monitoring Devices

192 CHAPTER 9 Calibration of CGM systems

Performance assessment
The performance of the calibration algorithm is assessed by comparing the CGM
calibrated profiles under test and the YSI laboratory references. First, the two
measurements are matched in time. Then, accuracy is quantified by computing
the following three metrics:
• Mean absolute relative difference (MARD) between the calibrated CGM profile
and the YSI measurements
• Percentage of accurate glucose estimates (PAGE), that is, the percentage of
estimated glucose values falling within either 20 mg/dL from the relative YSI
reference if YSI is lower than 80 mg/dL or within 20% of the relative YSI
reference if YSI is above 80 mg/dL.
• Percentage of CGM-YSI pairs lying in the “A” zone of the Clarke error grid
(CEG-A). The CEG is an error grid divided into five zones indicating the accuracy
of BG estimates generated by meters as compared to a reference value. In partic-
ular, zone “A” indicates accurate glucose results that do not lead to subsequent
wrong or dangerous treatments.
These three metrics are used to assess the accuracy of both the CGM profiles as
originally calibrated by the manufacturer and the CGM profiles as calibrated by the
Bayesian algorithm discussed here.
For each CGM profile under test, a subject-level analysis is performed by
computing the performance metrics for each dataset and then taking the population
statistics. The population performance indexes are obtained by computing the mean
and standard deviation of the metrics obtained in each dataset for normally distrib-
uted metrics and the median and interquartile range of the metrics obtained in each
dataset for nonnormally distributed metrics. Normality is assessed for each metric by
the Lilliefors test.
The statistical significance of the differences in performance metrics obtained
with the new and the original calibration is determined by the Wilcoxon signed-
rank test, a nonparametric-paired statistical test on the median of the performance
metrics distributions. In particular, we tested the null hypothesis “the median differ-
ence between the paired values of the two groups is zero” with a significance level of
0.05 on the performance metrics distributions obtained with the original calibration
against the new algorithm (for different calibration frequencies).

Results
The calibrated profiles obtained with the three scenarios listed earlier and the
original manufacturer CGM output are shown, for a representative subject of the
database, in Fig. 9.9, where also YSI references (not used by the calibration proced-
ure) are reported. Top panel refers to the original manufacturer calibration
(performed twice per day), whereas the other panels to the Bayesian calibration
algorithm and, in particular, from top to bottom, to the one-every-day, one-every-
two-days, and one-every-four-days calibration scenarios. The calibrated profiles
obtained with the Bayesian calibration algorithm are, independently from the