An artificial pancreas is a medical Internet of Thingsbased system consisting of a continuous glucose monitor, an insulin pump, and a micro-controller. The use of artificial pancreas systems is becoming increasingly popular amongst patients with type 1 diabetes due to its effective ability to allow the patient better control of his/her own blood glucose levels compared to other more standard treatments. In this paper, the problem of missing sensor readings in the glucose monitor data is considered. How should the microcontroller (which adjusts the insulin pump based on monitor readings) react when the glucose monitor stops transmitting for an unpredictable period of time? A strategy that answers this question is called a failure mode strategy. In this paper, several potential failure mode strategies are explored in the context of simulation experiments. Results are presented showing that at least one effective and simple failure mode strategy (0.5 & LR<72) exists.
An artificial pancreas (AP) [
In this work, the problem of missing data in the CGM trace is
considered [
Gaps in real CGM traces occur for a variety of reasons. In most
cases, they are benign (e.g. the sensor is no longer accurate and needs
to be replaced) but there is also the possibility of deliberate malicious
interference (for example, [
As mentioned, a simple default strategy if connectivity to the
sensor is lost is to simply turn the insulin pump off or switch to a preset
basal rate. However, this is unsatisfactory for the reasons mentioned
above. An alternative idea is to have the controller replace the
missing CGM readings with estimates, and then use these for the insulin
dose calculations so that the decisions on insulin delivery rate can
still be made. Unfortunately, there is currently no effective method
for dealing with missing CGM readings in real-time [
Therefore in this paper approaches based on changing the behaviour of the insulin pump when the sensor is down are considered. It is shown via a set of experiments involving virtual patients and a state-of-the-art AP controller that viable alternative failure mode strategies that can replace the simplistic “pump off” strategy exist. Note that imputation of missing glucose values is not a focus here; instead we are concerning with how the pump should behave when the sensor is unavailable. 2
To perform the comparison of failure mode strategies, a virtual patient simulator along with a modern AP controller based on fuzzy logic was used. 2.1
The simulator utilised in this work is an open-source T1D patient
simulator Simglucose [
The virtual patient model used by the simulator describes the glucose/insulin subsystem and is modelled using compartments. Important processes in the virtual patient simulator include a model of the gastrointestinal tract, from which the rate of appearance of glucose in the glucose subsystem is determined during simulated meals; the insulin subsystem, which models both the rate of appearance of insulin, its rate of degradation, and its interactions with the liver and body tissues; and the production and storage of glucose in the liver and its utilisation in muscle and adipose tissue. Each virtual patient in the simulator is generated randomly from a joint probability distribution over tens of physiological parameters, and ten different adult patients are currently available in the simulator.
The simulator models meals as well, and the timing and size of meals in terms of carbohydrate (CHO) count are also randomly generated from an appropriate probability distribution.
Figure 1 shows a trace of the simulator’s output for one virtual patient. The figure shows CGM readings, actual BGL (not always identical due to sensor lag and error), and the insulin pump rate. The normoglycemia range (see Table 1) is also shown. 2.2
The controller component of an AP serves the function of monitoring (and aggregating) incoming CGM sensor trace data, and then using this information to make making insulin pump rate adjustment decisions in real time (for example, to reduce the risk of hypoglycemia induced by a rapid decline in BGL).
Many algorithms have been proposed for solving this control
problem, and they can be broadly grouped into control theory-based
approaches such as proportional-integral-derivative (PID) controllers;
logic-based approaches; adaptive statistical based-approaches (e.g.
based on moving averages); and machine learning-based
approaches [
In order to circumvent these issues (mostly), we implemented the
fuzzy logic controller (FLC) proposed by Mauseth et al. [
These three inputs are then “fuzzified” (mapped to fuzzy sets), and following that a fuzzy lookup table containing the expert knowledge is consulted for optimal dose calculation. In general, the table is designed such that higher accelerations and velocities lead to higher doses, while low absolute BGL values and strongly negative velocities switch the insulin pump off.
The FLC does have a single patient-specific parameter, the “personalisation factor” (PF), which can range from zero to ten. The PF dictates the aggressiveness of the insulin treatment, with a value of ten corresponding to a very weak treatment (all doses being multiplied by a factor of 0.002), while a PF value of zero is the strongest treatment (all doses scaled by 1.74). Most patients should be expected to have a PF of around five, indicating a scaling factor of 0.92 of the dose computed by the FLC. 3
In this section, the experimental setup, in particular the methods by which the PF values are set on a patient-wise basis, and how artificial missing data gaps were generated in the simulated CGM trace are described. 3.1
Each individual has significantly different insulin sensitivity. Some are more sensitive to insulin, so their blood glucose levels drop rapidly in response to the same insulin dose than those who are less sensitive. Thus, the FLC’s PF value is necessarily unique for each individual.
To determine the optimal PF value for each adult virtual patient in the T1D simulator, the following procedure was followed: for each possible PF value, the simulator was run for ten virtual days at a sampling interval of five minutes. The amount of time spent in normoglycemia for each patient and each PF value respectively was recorded. The PF value for each adult virtual patient which maximised the time spent in normoglycemia was then selected. 3.2
One guideline concerning proper usage of CGM [
In this section, five different failure model strategies are described. As a baseline, results for an “ideal” set of simulations in which there are no sensor trace gaps are also included. The failure mode strategies are as follows:
This strategy simply switches the insulin pump off when sensor gaps are detected. 2.
is defined as the average value of the total insulin dose from the previous one hour. Since there are twelve readings per hour in an AP with a five minute reading interval, the average dose is defined as = 1 12 n=1 12 X di (1) where di is the insulin dose i adjustments previously. The strategy operates the pump at a constant rate of . 3. 0.5
On the basis that may produce doses that are sometimes too high, the 0.5 strategy is calculated in the same way, except that the constant dose size is halved.
Random choice constructs a method that randomly chooses a value from the last one hour of insulin dosages, on the basis that varying the dose size might be beneficial compared to the strategies which computes constant doses. The expected value of the random choice strategy equivalent to the dose as calculated by . 5. Hybrid method (0.5 & LR<72)
Following initial results showing that the 0.5 strategy was effective, a new strategy was defined that extended 0.5 with a constraint: if the last reading (LR) before the sensor is down is smaller than a preset value (in this case, 72.0 mg/dL), then the pump will be turned off; otherwise, the 0.5 strategy will be applied. 72 mg/dL was chosen as a threshold because it is just above the L1 hypoglycemia threshold defined in Table 1, and it is desirable that this threshold is not crossed. The equation for this strategy is: dose = (0 0:5 if LR <72.0 otherwise (2) 4
In each experiments, 100 simulated days per adult virtual patient per failure mode strategy was run. This gave a total of 10 6 runs per patient. Each run took approximately ten minutes on a laptop with a 2.6GHz Intel Core i7-9750 CPU processor. For all experiments, the simulated GuardianRT CGM was used in conjunction with the Insulet insulin pump.
Table 2 shows for each patient with their personalised PF value according to the method of selecting the optimal PF value described in the previous section.
The effectiveness of each failure mode strategy was then determined by calculating the time spent in standard glycemic ranges as defined in Table 1. As mentioned, hypoglycemia can be considerably more dangerous than hyperglycemia, so therefore time spent in the L2 Hypoglycemia range is the most significant statistic in the results.
Figure 2 shows that these five strategies provide similar performance in the L1 Hyperglycemia range. All the median values are
L2 Hypoglycemia L1 Hypoglycemia Normoglycemia L1 Hyperglycemia L2 Hyperglycemia
<54 mg/dL (<3.0 mmol/L) 54 to <70 mg/dL (3.0 to <3.9 mmol/L) 70 to 180 mg/dL (3.9 to 10.0 mmol/L) >180 to 250 mg/dL (>10.0 to 13.9 mmol/L) >250 mg/dL (>13.9 mmol/L) quite close to each other (i.e. between 18.21 and 19.44). and Rand choice perform quite well most of the time. The lowest percentage obtained due to Rand choice is below 5% for one patient. However, 0.5 and hybrid strategy have a more stable performance r e p y1H 51 L e m it % 01 5 2 0 2 5 0
19.44 18.4 18.51 19.2
19.19 18.21 with no significant outliers.
% time L2 Hyper in different strategies nificantly underperform compared to the other strategies. In terms of the more dangerous L2 hypoglycemia range (Figure 6), the pump off and hybrid strategies work best.
% time L1 Hypo in different strategies ypo 3 H 1 L e m it% 2 o p yL2H .06 e m i t% .40 5 4 1 0 5 0 0 8 0 2 0
No MV
Figure 4 depicts time spent in normoglycemia. Unlike the other figures where lower is better, in this case higher is better. The figure indicates that the last four strategies output similar median values which are all quite close to the baseline no missing value strategy. The best two strategies that give us the highest median time are again and Rand choice. However, it is quite noticeable that these two methods, especially Rand choice, as not stable, as observed before. The percentage time spent in normoglycemia can drop to under 10% for one of these strategies. The 0.5 & LR<72 strategy gives the most stable performance across virtual patients.
Figure 5 and 6 show time spent in the hypoglycemic ranges. From these box plots, it can be observed that both and Rand choice sigIn this study, it was found that the hybrid method works well and is furthermore simple to implement. In most cases, it was close to the no missing values baseline strategy, especially in terms of time spent in the critical L2 hypoglycemia range.
The hybrid method could therefore be used as a baseline for the development of more sophisticated failure mode strategies (such as those based on online CGM imputation or machine learning methods) in the future. It is also important to test this failure mode strategy with other controllers and on other twenty patients (i.e., ten children and ten adolescent patients) in addition to the FLC and ten adult patients analysed in this paper.