Page 180 - Glucose Monitoring Devices

182 CHAPTER 9 Calibration of CGM systems

Two approaches to optimize the computational complexity
Most of the algorithms proposed for improving CGM performance employ sophis-
ticated models and signal processing features that, although still allowing the imple-
mentation on wearable devices/smartphones, increase the computational complexity
and processing delay compared to the simple linear regression techniques. With the
aim of reducing the delay due to signal processing, Mahmoudi et al. proposed a
multistep calibration algorithm based on rate-limiting filtering, selective smoothing,
and robust regression [37]. The rate-limiting filter limits the rate of change if a phys-
iological threshold is exceeded; the selective smoothing is applied if the signal is
noisy, that is, if the number of zero crossings of the signal first-order differences
exceeds a predefined threshold; the robust regression then converts the raw measured
current to BG levels using reference SMBG measurements (for a maximum of four
references per day). The application of the filtering step to only the noisy parts of the
signal lowered the delay introduced by the signal processing of the CGM profile.
Another approach that has low computational complexity as a major strength
was proposed by Kirchsteiger and colleagues employing linear matrix inequalities
techniques, resulting in convex optimization problems of low complexity [38,39].
The authors proposed two different parametric descriptions of the relationship
between IG and BG and a constructive algorithm to adaptively estimate the unknown
parameters. The algorithm explicitly considers the measurement uncertainty of the
device used to collect the calibration measurements, which was first pointed out by
Choleau and colleagues [18]. Moreover, the algorithm embeds an automatic feature
to detect fingerprick measurements, which are not suitable to be used for calibration.

Deconvolution-based Bayesian approach
The uncertainty in the reference SMBG samples used for calibration is a key issue in
the development of robust calibration algorithms. The real-time deconvolution-
based approach proposed by Guerra et al. [40] demonstrated its robustness against
both temporal misplacement of the SMBG references and uncertainty in the BG-
to-IG kinetics model. The authors proposed a real-time signal-enhancement module
to be applied to the CGM sensor output to improve the accuracy of the device. The
algorithm compensates the distortion due to the BG-to-IG dynamic by means of
regularized deconvolution [41] and relies on a linear regression model that is
updated each time a pair of SMBG references is collected. Significant accuracy
improvements were observed both on simulated and real datasets. The
deconvolution-based approach of [40] was further developed in Refs. [42,43], where
it was directly applied to the raw measured signal rather than in cascade to the CGM
sensor output. The algorithm fits the raw current signal against BG references
(collected twice a day) using a time-varying linear calibration function whose
parameters are identified in the Bayesian framework using a priori knowledge on
their statistical distribution. The BG-to-IG kinetics is compensated, as in
Ref. [40], via nonparametric deconvolution. Results showed significant accuracy
improvements compared to the manufacturer calibration.