Page 233 - Glucose Monitoring Devices

236 CHAPTER 11 Retrofitting CGM traces

With regard to CGM data, seldom they suffer isolate readings affected by a
significantly larger error with respect to usual one, commonly referred as spikes
and illustrated in Fig. A.1, at about 15:00. Moreover, as documented in Ref. [29],
CGM may suffer failures related to the biomechanics of the sensor-tissue interface:
pressure on the sensor site and foreign body response can create transient effects
which generate unphysiological changes of the CGM output, called loss of sensi-
tivity (Fig. A.1, at about 10:30, for 20 min).
The data preprocessing step of the retrofitting method aims to detect and isolate
possibly erroneous data points. For most of these points, automatic discharge/correc-
tion can be performed, but sometimes human visual inspection and intervention is
required. In the second case, the operator running the off-line analysis can manually
correct, confirm, or exclude the isolated data. The data preprocessing step takes as
inputs row CGM and BG data and returns as output outliers-checked data.
In this work, we limit ourselves to simple fault detection heuristic, that checks
data consistency with physiological constraints: for instance, a rate of change larger
than 5 mg/(dL$min) or any values outside the range ½40e400Š mg/dL are considered
suspicious and isolated. The physiology-inspired rules are reported in detail below:
• Any reference BG or CGM value in outside the range [40e400] mg/dL is
prompted for confirmation to the operator running the retrospective analysis.
• If the Dexcom SEVEN PLUS sensor is used, given that it only produces CGM
values in the range [40e400] mg/dL, any eventual value outside this range has
to be interpreted as an internal error code and therefore automatically excluded.
• A conservative bound on BG rate of change, r BG , is checked:

h i
BGðt iþ1 Þ BGðt i Þ mg
5 (11.23)
jr BG ðt iþ1 Þj ¼

t iþ1 t i dL$min
• A similar control is performed on the CGM rate of change, r CGM , but the
threshold is loosened not to exclude poorly calibrated data portions:

h i
CGMðt iþ1 Þ CGMðt i Þ mg
10 (11.24)

jr CGM ðt iþ1 Þj ¼

t iþ1 t i dL $min
• Possibly unreliable reference data (see Fig. 11.2, top panel) are abnormally low
(high) with respect to previous and future references. This situation is detected
checking the following condition:
mg mg
h i h i
jr BG ðt iþ1 Þj 1 and jr BG ðt iþ2 Þj 1 (11.25)
dL$min dL$min
• but

(11.26)
signðr BG ðt iþ1 ÞÞssignðr BG ðt iþ2 ÞÞ