Page 61 - Glucose Monitoring Devices

Accuracy and its consequences 59

variability and are applicable only to CSII-based therapy. For example, in Ref. [18],
all patients are assumed to behave randomly according to a pattern consistent with
observational study data. Behavior is limited to the likelihood of fingersticking
under circumstances that arise in CSII treatment such as hypo and hyperglycemia
awareness, fingersticking during meals, and fingersticking 2 h after a meal. Other
behaviors, such as meal-related ones are assumed to be independent [39,75].

Integrated metabolic/behavioral simulation
An alternative simulation approach [77] uses a hybrid approach where compart-
mental estimates are complemented with data-driven (patient-specific) quantifica-
tions of metabolic and behavioral variations obtained by model inversion. These
models have been extensively used in the development, evaluation, and risk
assessment of insulin dosing technologies [18,48,78].

Modeling glucose monitoring devices (device and lot variability)
The measurement error of several glucose monitoring systems has been modeled
extensively [35,79,80]. In particular, models for SMBG meters became available
recently [18,48,81] in great measure due to the availability of and extensive array
of accuracy assessments [8,10,11,13,14]. Meter and CGM models make it possible
to realistically simulate not only many commercially available monitoring devices
[69] but also meters/sensors with hypothetical characteristics as in Refs. [48,78]
where meters/sensors with increasing noise are systematically simulated. When
assessing a particular meter brand, it is important to understand variability across
meters of the same brand and model, or even across test-strip lots within the same
brand [27]. This can be achieved by explicitly characterizing the distribution of
model parameters [79,81]. Alternatively, a separate model can be identified for
each meter dataset. During a simulation, the collection of models obtained this
way can be sampled randomly [18,76,78,80].

In silico accuracy studies
An in silico accuracy study puts together all the elements described earlier.
As shown in Fig. 4.4, a highly accurate glucoseeinsulineglucagon metabolism
model can be integrated with a meter model and a behavioral model of the patient,
to understand the effect of dosing decisions. This approach was originally developed
to test the performance of closed-loop algorithms [60,71,72]. Building on these
studies [42,48,76], simulation has been used to quantify SMBG accuracy error
effects on glucose variability, risk of hypoglycemia, and average glucose. As
described in Fig. 4.4, these studies simulate virtual patients making decisions
according to a set of prespecified decision-making rules informed by their behav-
ioral model as well as simulated SMBG meter measurements. The decision is fed
back to the simulator to estimate metabolic dynamics until the next decision point
in time. Information about glucose concentration, meter readings, meals, insulin
bolus, and rescue carbohydrates are recorded throughout the simulated scenario.