Page 63 - Glucose Monitoring Devices

Accuracy and its consequences 61

characterize the relationships between accuracy characteristics and clinical
outcomes. For example, in Ref. [78], simulation is used to determine mean absolute
relative difference (MARD) levels that are acceptable if a CGM meter is used non-
adjunctively (e.g., MARD 10%). Similarly, simulation results have been used to
understand accuracy requirements to achieve clinical performance requirements
[30]. A recent study more relevant to glucose meters [76] shows that a regression
model can be used to characterize relationships between SMBG bias and noise
and clinical outcomes such as severe hypoglycemia events and HbA1c, total daily
insulin, and fingerstick use.

Long-term health and complications
From in silico results to long-term complications
Relating quality of glycemic control and long-term clinical complications is the last
step of the process, and perhaps the most difficult to perform [84].
One possible approach is to use population-level published literature relating
changes in a clinical outcome metric to the incidence of long-term consequences.
For example, in Ref. [54] authors relate changes in HbA1c baseline values to the
incidence of complications such as retinopathy, renal failure, heart disease, and their
associated costs. Fig. 4.5 shows an example of such an approach. Changes in HbA1c
(x-axis) have a corresponding increase in the incidence of diabetes-related compli-
cations (y-axis). Knowing the number of events before and after the change allows us
to calculate the effect on the total complication-related cost. By filtering the under-
lying literature for, for example, geographical region, patient age, diabetes type, or
duration of the disease, the approach can be adapted to nonaverage populations.
An alternative approach [85] used the results of a health survey (EQ-5D) to
develop a linear regression model associating changes in HbA1c to a measure of
health (or health utility).
A third approach, used by Refs. [85e87], is to relate changes to HbA1c with
changes to the progression of diabetes complications. Here, a Markov cohort
modeling approach is used [88], as illustrated in Fig. 4.6. In this approach, each
patient is considered to be in a state of progression, expressed by the state of the
Markov chain. The transition from one state to another marks the progression of
the disease and is expressed in conditional probabilities or transition rates. By
analyzing the dynamics of the Markov chain, it is possible to assess the population
distribution across disease progression states. Examples of this modeling approach
are the Michigan Model for Diabetes [88], ECHO-T2DM [89], and the CORE model
[90,91]. In Refs. [86,87], the transition probabilities in the chain are also conditional
on HbA1c (illustrated as dotted lines in Fig. 4.6). For example, it is assumed that the
probability of progressing to full blindness in the next period (see the bold line in
Fig. 4.6) is conditional on the current state of the patient (Macular Degeneration)
and its HbA1c level.