Page 194 - Glucose Monitoring Devices

196 CHAPTER 9 Calibration of CGM systems

the performance metrics obtained with the original calibration, the Bayesian algo-
rithm with one calibration per day, one calibration every 2 days and one calibration
every 4 days. Columns 6e8 report the corresponding p-values by the Wilcoxon
signed-rank test. The indexes reported in Table 9.1 show that the Bayesian calibra-
tion algorithm improves sensor accuracy compared to manufacturer calibration,
independently from the frequency of calibrations. The improvement achieved
with the Bayesian algorithm is statistically significant (P <.05) for all the
considered metrics and for all the calibration frequencies tested. In addition, no
statistically significant difference (P >.05) is found between the performance
metrics distributions obtained with the Bayesian algorithm by using different
calibration frequencies, that is, one per day versus one every 2 days versus one every
4 days (P not shown).
It is interesting to consider also an extreme scenario where no calibration, except
the initial one, is performed. In this zero calibrations scenario, the calibration param-
eters are estimated about 2 h after sensor insertion and used for the entire monitoring
session without any further update. As the estimation is performed in the Bayesian
setting exploiting only a pair of SMBG samples, it strongly relies on prior informa-
tion. In this case, the following median values are obtained: 12.03% MARD, 85.67%
PAGE, and 85.62% CEG-A. Although these indexes are still slightly better than
those of the original manufacturer calibration (see Table 9.1), no statistically
significant differences are observed. Indeed, the larger the number of calibration
references (as moving from day 1 to day 7), the more the posterior distributions
differentiate from the priors. This phenomenon is depicted in Fig. 9.11, where the
posterior distributions of the three calibration parameters are reported, centered
with respect to the a priori expected values. We can observe that, not surprisingly,
as moving from day 1 to day 7, that is, as the number of calibration references
increases, the posterior distributions differentiate from the priors (this is more
evident for parameters b and s 2 ). Notably, the prior distributions are quite flat,
much more than the posterior distributions. This suggests that the information
brought into the Bayesian estimation process by the calibration references plays a
key role in determining the parameter values and that the use of only one calibration
reference would let the estimate to excessively rely on prior information.

Conclusions
Most of the commercially available CGM sensors need to be calibrated to convert
the raw measurements to glucose values. To preserve sensor accuracy, manufacturer
instructions recommend a calibration at least every 12 h. Simple linear regression
techniques have been extensively employed for calibration since the commercializa-
tion of the first CGM devices. Although their simplicity and ease of implementation
in wearable devices represent the fundamental strength of these approaches, sensor
inaccuracy problems and the need for frequent recalibrations called for the develop-
ment of more sophisticated techniques.