Page 224 - Glucose Monitoring Devices

Retrofitting outpatient study data 227

Finally, the constrained Tikhonov regularization problem has the form:
r
argmin (11.21)
bb g ¼ J r ðbgÞ
r
l ci ðbgÞ C$bg u ci ðbgÞ
where
r r T 1 r r T T d r
reg %ðbg Þ F F bg
J r ðbg Þ¼ðcgm recal Gbg Þ S ðcgm recal Gbg Þþ g
(11.22)
First addend in (11.22) penalizes inadequate data description, while the second
addend takes into account the physiological prior knowledge of blood glucose
is the regularization parameter that trades
smoothness, where d is fixed, d ¼ 2. g reg
off data fit and smoothness of the resulting profile.
The current implementation employed a fixed g reg , manually tuned to achieve
satisfactory performances on a validation dataset. The implementation of an
from the data is deferred to future works.
reg
algorithm that allows learning of g

Retrofitting outpatient study data

In this section, we show that the retrofitting algorithm can be used to enhance
precision and accuracy in CGM data collected during outpatient clinical studies
such as [11e14], i.e., in a setup offering a relatively large number of highly accurate
references to retrospectively enhance the CGM.
An in-depth discussion of this setup can be found in Ref. [6].
We start from a dataset, called “original dataset,” collected in an inpatient study
and offering frequent reference BG measurements. References BG are then divided
into training-set references, available to the retrofitting algorithm, and test-set refer-
ences. The training set will be called “outpatient-like dataset.”

Original dataset
The data used in this section were collected during a large multicenter inpatient clin-
ical trial[20], conducted within the EU-funded project AP@home [21]. The trial
aimed to compare two different closed-loop algorithms against the standard
open-loop therapy and involved 47 patients in six European centers. Each patient
underwent three admissions, lasting about 24 h and employing three different
therapies, i.e., open-loop (OL) and two different closed-loop algorithms (CL).
Frequent BGs were collected throughout the admission, every hour during the night
and at least every 30 min during the day, resulting in the availability of w55 BG
references/day. BG references were measured with YSI2300 STAT Plus analyzer
(YSI, Lynchford House, Franborough, United Kingdom) and the CGM sensor was
the Dexcom SEVEN PLUS CGM sensor (Dexcom Inc., San Diego, CA, USA).
More details on the trial can be found in Ref. [20].