Page 85 - Glucose Monitoring Devices

84 CHAPTER 5 Modeling the SMBG measurement error

error in the low glucose range (BG < 100 mg/dL) and one to describe the relative
error in the high glucose range (BG 100 mg/dL). The methodology by Vettoretti
et al. was applied to two literature datasets of SMBG data, collected by One Touch
Ultra 2 and Bayer Contour Next USB (Bayer HealthCare LLC, Diabetes Care,
Whippany, NJ), and reference BG samples, measured by the YSI (YSI Inc., Yellow
Spring, OH) laboratory equipment [40,41]. A two-zone SMBG error model employ-
ing skew-normal distributions was derived for each of the two considered datasets
[40]. The models were validated by goodness-of-fit tests, which showed the superi-
ority of the two-zone skew-normal PDF model compared to the simple Gaussian
PDF model previously employed in the literature.
In the following section, the method proposed by Vettoretti et al., which currently
represents the state-of-the-art approach for modeling the SMBG error distribution, is
described in detail.

The state-of-the-art modeling method by Vettoretti et al.
To create a model of the SMBG measurement error PDF, two strategies are possible.
The first is to resort to a parametric approach in which it is assumed that the PDF we
want to describe presents a known shape characterized by a fixed, and usually small,
number of parameters. The disadvantage of this approach is that it requires some
assumptions. The second possibility is to use nonparametric approaches, but coping
with the risk of overfitting (i.e., to include dataset-specific phenomena in the model)
can be difficult. Moreover, nonparametric methods do not allow to obtain any infor-
mation on the PDF’s properties, for example, they do not allow to establish if the
distribution is unimodal, it presents a positive skewness, etc.
Recently, our research group at the University of Padova proposed a new meth-
odology to develop and validate parametric models of the SMBG error PDF [40],
which consists of four steps:
A. definition of a training and a test set;
B. identification of BG zones of the training set in which the errors present a
constant-SD distribution;
C. maximum-likelihood (ML) fitting of a PDF model to errors in each identified
zone of the training set;
D. model validation by comparing the distribution of random samples simulated by
the model with the distribution of data in the test set.
In the following sections, these four steps are described in detail. Of note, the
method by Vettoretti et al. is general and can be easily applied to any dataset
containing SMBG datapoints and matched BG references. Later in this chapter,
the method will be applied to two case studies to develop a model of SMBG
measurement error for two commercially available glucose meters (One Touch Ultra
2 and Bayer Contour Next USB).