Page 179 - Glucose Monitoring Devices

State-of-art calibration algorithms and today’s challenges 181

for BG-to-IG kinetics. A further development, with a direct application to the cali-
bration problem and incorporation of the BG-to-IG dynamic model, was given by
Facchinetti et al. [32]. The authors proposed an extended Kalman filter method
that works in cascade to the standard device calibration to enhance sensor accuracy.
By taking into account BG-to-IG kinetics, using a model to describe the variability
of sensor sensitivity, and exploiting four BG reference samples per day, the method
significantly improves CGM accuracy when applied to synthetic data. However, its
real-time implementation is not straightforward, requiring the knowledge of the
variances of both state and measurement error processes, as well as an initial
burn-in interval.

Methods relying on autoregressive models
Another approach for real-time glucose estimation based on autoregressive (AR)
models was proposed by Leal et al. [33]. The study used AR models to estimate
BG from raw CGM measurements. Data acquired from 18 T1D patients were
used to train a population AR model, which was then incorporated into a calibration
algorithm for real-time BG estimation. The raw sensor signal, used as the indepen-
dent variable, and the BG concentration, considered as a dependent variable, were
both normalized based on the maximum range of the available signals. The best
overall estimated model, with a third-order BoxeJenkins structure and fixed
parameters, enhanced CGM performance, especially in hypoglycemia detection.
Significant improvement in hypoglycemia detection was also obtained by the
same authors in another study performed on 21 patients with T1D where a new linear
regression algorithm with enhanced offset estimation was proposed [34].

A calibration method integrating several local dynamics models
Barcelo-Rico and colleagues proposed an alternative calibration algorithm based on
a dynamic global model of the relationship between BG and interstitial CGM signal
[35]. The algorithm integrates several local dynamic models, each one representing
a different metabolic condition and/or sensor-subject interaction. The local models
are then weighted and added to compose the global calibration model. Inputs of the
model are the signal measured by the sensor and other signals containing informa-
tion relevant to glucose dynamics, which are normalized in magnitude using
population parameters. The algorithm showed improvements in CGM sensor
accuracy, although it was tested on only eight healthy subjects and a more extensive
assessment on the diabetic population would be needed to confirm the findings.
Further development of the algorithm was proposed in Ref. [36], where an adaptive
scheme is used to estimate a patient’s normalization parameters in real-time instead
of using simple population parameters. The results on 30 virtual patients showed that
the adaptation of normalization parameters further improved the performance of the
algorithm, as they were able to compensate for sensor sensitivity variations.