Page 200 - Glucose Monitoring Devices

CHAPTER

CGM filtering and

denoising techniques 10

Andrea Facchinetti, PhD, Giovanni Sparacino, PhD, Claudio Cobelli, PhD
Department of Information Engineering, University of Padova, Padova, Italy

Introduction
Continuous glucose monitoring (CGM) sensor data can be affected by several sour-
ces of error, for example, due to imperfect calibration, sensor physics, chemistry, and
electronics, which can affect both accuracy and precision of CGM readings [1e6].
In particular, the CGM signal is also corrupted by a random noise component, which
dominates the true signal at high frequency [7e9].
The presence of the random noise component on CGM data is evident from
Fig. 10.1, which illustrates two representative time series (black line) measured in
two diabetic subjects through a commercial CGM device, the Glucoday (Menarini,
Firenze, Italy), a minimally invasive microdialysis sensor that provides glucose
readings every 3 min (data taken from Ref. [10], where details on the sensor can
be found). On the top panel, 1-day CGM data of the first representative subject
shows are clearly corrupted by a large noise component. On the other hand, the
bottom panel shows 1-day CGM data of a second representative subject, in which
the noise variance seems, in general, smaller than for subject #10, even if large
spurious spikes (possibly due to patient movements that episodically may perturb
sensor behavior) are present.
Spurious spikes and oscillations can affect the performance of any algorithm
based on CGM data for therapeutic decisions and suggestions. For instance, the
generation of hypoglycemic and hyperglycemic alerts is strongly influenced by
the CGM sensor’s accuracy and precision, with a percentage of false alerts of the
order of 50% in the worst cases [11,12]. Another example in which the random
fluctuations around the actual CGM value could be critical is in the calculation of
insulin boluses, where it can lead to under/overestimations of the insulin amount
of the injected, with the possibility of dangerously increasing risks of hypo/hyper-
glycemia [13]. Finally, other CGM-based applications negatively affected by the
presence of random noise include glucose predictors [14e16] and closed-loop
control strategies [17,18]. In this chapter, we will deal with the problem of reducing
the impact of random noise on CGM data.

Glucose Monitoring Devices. https://doi.org/10.1016/B978-0-12-816714-4.00010-7 203
Copyright © 2020 Elsevier Inc. All rights reserved.