A Review of the Current Challenges Associated with the Development of an Artificial Pancreas by a Double Subcutaneous Approach

In parallel with the development of an insulin-suspend function, several groups have developed and refined associated algorithms, some of which are now commonly employed for the double subcutaneous AP approach.

One frequently used algorithm is the proportional-integral-derivative (PID) control method, which is commonly used in industrial applications. It takes into account the deviation from a target glucose value (the proportional term), the past values of the deviation (integral term), and the rate of change in the deviation (the derivative term) [36‐39].

The most commonly used algorithm is model predictive control (MPC), which predicts the system’s behavior and calculates the optimal insulin infusion (optimal with respect to, for example, the deviation from the target glucose level, the deviation from a target zone, the insulin expenditure, etc.). Some MPC algorithms also take into account factors such as meal/carbohydrate input, variable insulin sensitivity, and recent glucose excursions when tailoring the dose of fast-acting insulin [23, 27‐31, 35]. Most MPC algorithms aim to achieve a predefined glucose target, some aim to keep the glucose level within a particular range (“zone MPC”), and some do both.

Most recent studies lasting at least five consecutive days in the outpatient setting have explored overnight glucose control, and most of those studies showed significant improvements in the time spent within either a narrow or wider range of glucose levels (3.9–8.0 or 3.9–10.0 mmol/L, respectively), ranging from 50 to 85%, when using closed-loop control [21, 29‐35]; other studies found no difference from open-loop control [36, 37]. In addition, many of these studies showed that less time during the night was spent in hypoglycemia [21, 29‐31, 34], while others found that there was no difference in time spent in hypoglycemia from that obtained with open-loop control [30, 32, 33, 36].

In contrast to the many overnight studies that have been performed, only a few studies have explored outpatient closed-loop control during both the day and the night.

Whereas around 85% of the time was spent within the therapeutic range in some of the night-time studies [29, 32], the best achievement in the day-and-night studies was around 75–80% of the time spent in the therapeutic range [29, 35]. Among the seven day-and-night closed-loop control studies included in Table 1, four found that more time was spent within the therapeutic range [29‐31, 35], three found that less time was spent in hypoglycemia [29‐31], and three found that less time was spent in hyperglycemia [29, 30, 35] than when open-loop control was employed. The longest day-and-night study with closed-loop control lasted 12 weeks [30].

At present, several studies with even longer periods of outpatient double SC closed-loop control are ongoing. While there are many studies which show that double subcutaneous closed-loop control is better at keeping blood glucose within the therapeutic range and is safer (considering the time spent in hypo- and hyperglycemia) than traditional open-loop control, improved average glucose levels are not a uniform finding when double subcutaneous closed-loop control is applied.

“Latency” and similar terms that are used in the AP literature relate to several dynamic phenomena that are of fundamental importance in a closed-loop control system such as an AP. We therefore offer a brief and pragmatic explanation of the most prominent concepts—namely time delays and time constants—to aid readers who are not familiar with dynamic systems theory [42]. The applicability of time constants in the present context is justified by the fact that the human glucose metabolic system is predominantly linear within its normal operating range and can thus be well approximated by linear models [43].

A time delay τ (cf. Fig. 1b) is typically associated with the transportation of a substance over a certain distance (e.g., the transport of glucose or insulin through the circulatory system or through tissue), and it denotes the time that passes from the application of a stimulus to a system until the first trace of the incident is theoretically detectable in the system’s output.

A time constant T (Fig. 1c) is associated with, for example, the diffusion of a substance across a thin, semipermeable barrier separating two compartments, each of which has a uniform concentration of the substance (e.g., the diffusion of glucose or insulin across cell membranes, or across a protective membrane separating a glucose sensor from the surrounding fluid). A concentration gradient across the barrier will result in the net transport of the substance towards the region of lowest concentration, which gradually reduces the gradient until a dynamic equilibrium is reached. The time constant T associated with this process is the time from the initial establishment of the gradient until the concentration has reached 63% of its asymptotic level.

The effect of a time constant can in part be negated by appropriately filtering the output signal; this is limited only by the signal-to-noise ratio. In contrast, time delays cannot be negated in real time because this would require the ability to predict future events.

When it comes to closed-loop control, both time delays and time constants limit the obtainable bandwidth (i.e., the usable frequency range) of a closed-loop system. In an AP system exhibiting multiple time constants and/or delays (Fig. 1d) in the path from infused insulin to measured glucose, it is the sum of all the time delays and the larger time constant(s) that limit how quickly the AP can respond in order to correct glucose level deviations; larger latencies inevitably lead to larger glucose level fluctuations. This fact is at the core of this paper’s message.

In a healthy nondiabetic person, insulin is released from the β cells in the pancreas 2–3 min after the first increase in circulating glucose [44, 45]. This first phase of insulin response consists of the release of stored insulin and proinsulin from granules within the β cells. After about 10 min, a second and much stronger phase of insulin release commences. This phase consists of freshly synthesized insulin in response to the increased glucose levels sensed by the β cells, and lasts until the glucose levels normalize. At the end of this process, meal-initiated insulin secretion is shut off and insulin secretion is kept at a low level to maintain a steady glucose level. This biphasic pattern is only seen during fast changes in glucose levels as achieved by intravenous (IV) glucose infusions; it is absent during oral glucose loads, where only the second phase of insulin release is observed [45‐47].

In the fasting state, insulin is typically released in bursts 4–5 min apart, and the half-life of insulin in the blood is about 11 min. During the fasting state, approximately 70–75% of the insulin is secreted in a pulsatory manner in a canine model [48, 49]. Between these bursts, there is little or no release of insulin from the β cells. It is not known how this release of insulin from β cells is synchronized. However, this pulsatile pattern may be an important factor in the effect of insulin at the cellular level, as has been shown for other hormones with pulsatile release. The pulsatile delivery of insulin to the liver may also be important for increasing hepatic insulin sensitivity and achieving the full effect of insulin [48, 50, 51]. In healthy lean subjects without diabetes, approximately 50% of the insulin appearing in the peripheral circulation is delivered during the basal state in-between meals [52].

Insulin from the pancreatic β cells is released to the portal vein and arrives at the liver at a high concentration. During the passage through the liver, much of the insulin is absorbed in what is called the first-passage effect. Recent data indicate that as much as 80% of the insulin delivered in the bursts is absorbed, while as little as 20% is absorbed between bursts [53, 54]. This means that the exposure of the hepatocytes to insulin probably oscillates to a significantly greater degree than is reflected in the insulin-level oscillations in the portal and systemic circulation. The significance of this phenomenon is unknown. However, it may be important to the inhibition of hepatic glycogenolysis, the stimulation of glucose uptake from the blood arriving from the gastrointestinal tract, and the inhibition of hepatic gluconeogenesis. It has been suggested that the liver adapts its ability to withhold insulin to the levels in the portal circulation and the levels needed to suppress glucose in the systemic circulation [55‐57].

After an oral glucose load or a normal meal in a healthy, lean, physically active person, both glucose and insulin will return to their starting levels within 120–150 min [58]. However, in overweight or obese subjects, it may take longer to return to the normal values [52]. In healthy subjects, this means that glucose and insulin will be at, or close to, their steady-state levels during shorter daytime periods and most of the night [52].

One of the important counterregulatory mechanisms to avoid hypoglycemia is the release of glucagon from pancreatic α cells. This mechanism is blunted in patients with long-standing DM1, and although the α cells retain their ability to produce glucagon, the factors inducing the release of glucagon during hypoglycemia seem to be absent [59, 60].

To our knowledge, all commercially available glucose monitoring systems measure glucose in SC tissue using enzymes belonging to the oxidoreductase family. The oxidation of glucose is followed by an electrochemical or optical assay procedure to derive the glucose concentration by measuring the enzymatic products using an amperometric or a chemoluminescent technique, depending on the assay [61, 62].

There is a physiologic latency from the point in time when a certain glucose level is present in the blood until the same level is reached in the interstitial fluid (ISF) in SC tissue [63‐66] (Fig. 2). This latency is at least 6–7 min, but the physiologic latency between blood and ISF in SC tissue may be 15 min or more in some patients with DM1 [67‐69]. Further, there is normally sensor latency, which is the latency between the appearance of glucose in the ISF surrounding the sensor and its consequent appearance in the sensor signal (Fig. 2). This sensor latency may consist of a reaction time (e.g., if glucose needs to react with enzymes or other substances in a typical amperometric sensor), a diffusion time (e.g., the time taken to fill a probe volume in a spectroscopic sensor), or other sources of latency such as the computation time of the sensor’s software. In addition, commercially available SC CGM devices do not perform truly continuous glucose measurements [70]. Most devices provide an average of the glucose level over the last 5 min once every 5 min. This averaging procedure is commonly used to reduce the effect of short-term random noise (which is due to both intrinsic properties of the sensor and extrinsic random factors originating from the sensor’s chemical and physical environment) as well as to save battery capacity. Slowly varying signal variations due to, for example, biofouling are typically taken care of by repeatedly calibrating the sensor during its lifetime. Nevertheless, some recent CGM technology does not require frequent capillary calibrations [71].

The overall latency from physiologic and sensor-based sources may range from at least 8–10 min to above 20 min, which is far longer than the timescale required for real-time sensing of circulating glucose excursions experienced by healthy β cells. Some of these dynamics may be compensated for by the algorithm used to transform the raw sensor signal into an estimated glucose concentration. However, none of the companies in the CGM field publish their algorithms, so an external independent evaluation of the potential and limitations of this approach is not possible. The precision and accuracy of CGM have steadily improved, but such measurements still present limitations [72]. Although certain anthropometric measures (height, weight, skin thickness) seem to have little influence on accuracy, the impact of the foreign body response potentially limits the longevity of transcutaneous sensors [62, 70]. Further, although improved in the latest versions of CGM devices, the local ISF glucose level and thus the observed SC CGM sensor accuracy relative to blood glucose may be significantly influenced by local effects such as skin temperature, movements, the foreign body reaction, and pressure [73]. SC CGM devices typically show reduced accuracy for the first couple of days after insertion [62, 74]. This limitation should not be overlooked for a sensor with a usable lifetime of seven days. Certain drugs are also known to interfere with the measurements; for instance, the commonly used drug acetaminophen disturbs the enzyme-based CGM such that it leads to the overestimation of glucose levels in the ISF [75‐77]. In addition, the mean average relative difference (MARD) between the measured and true glucose levels will be larger during periods of rapidly changing blood glucose than during periods of more slowly changing glucose levels [78]. An important safety issue is that current CGMs show their lowest concordance when capillary glucose is measured in the lower or hypoglycemic range. Although evaluations of the hypoglycemic range are often missing from reports evaluating CGM, there are examples which demonstrate that deviations from capillary measurements are highest when measurements are in the hypoglycemic range as compared to other ranges [79‐82]. In other words, improvements to CGM are needed before it truly reflects blood glucose levels in a robust and reliable way.

One aspect of SC CGM that has hardly been discussed is the inflammatory reaction that is initiated whenever a foreign body is introduced into SC tissue [62, 83]. Patients experience this as a drift in sensor readings during the period in which they wear the sensor, and this partly explains why sensors need to be calibrated regularly. After decades of SC insulin infusion, most patients with DM1 have varying degrees of SC fibrous tissue formation at insulin injection sites. The fibrous-tissue-initiating effect of decades of SC CGM use adds to the fibrous tissue formation induced by SC insulin infusion. Most patients that are using SC CGM have used it for less than a decade. Accordingly, one might expect less predictable SC insulin absorption in patients with DM1 when they have been using this new technology in combination with SC insulin infusion for decades.

One challenge when designing an AP is the need for redundancy in glucose sensing. Ideally, this should be achieved by sensing glucose via two completely different techniques in two different tissues. Some AP groups have introduced redundancy by using two identical CGMs concomitantly [22, 23, 26‐28]. A downside of this is that it is a weak kind of redundancy and a probable cause of increasing long-term fibrous tissue formation.

State of the art insulin delivery in patients with DM1 is continuous injection from an insulin pump via a flexible plastic cannula inserted into subcutaneous tissue [84] (“CSII”). The insertion of a foreign body into subcutaneous tissue leads to a foreign body reaction. This effect can be reduced if the cannula is substituted and inserted at a different site after a few days. Local skin complaints (pain, lipohypertrophy) at insertion sites are quite common among insulin-treated patients, where lipohypertrophy is a consequence of the ability of insulin to stimulate local growth. A more important effect is that the pharmacokinetics of insulin uptake are probably altered in these patients, although the results of the few studies that address this issue are conflicting [85]. In the clinical setting, patients who have had DM1 for decades often report that their meal bolus expires earlier, or the opposite—that it has a delaying effect [86].

Importantly, fibrous tissue formation may cause varying and unpredictable absorption of insulin, and it is a major cause of insufficient blood glucose control and unpredictable hypoglycemic episodes. Varying insulin absorption can also occur independent of any foreign body reaction or fibrous tissue formation [86]. After a single injection of the fastest-acting “meal” insulin, the maximum insulin concentration in the blood is reached after 45 min, and the insulin level in the blood is still markedly increased after 3 h [87]. More importantly, the glucose-lowering effect is at its maximum between 90 and 120 min after injection, and this effect is still present after more than 5 h.

Although rapid-acting insulin yields more stable results when delivered by CSII as compared to intermittent SC bolus deposits, considerable day-to-day variation is still observed [88]. This intra-individual effect may vary from hour to hour, and may depend on the local skin temperature (which affects subcutaneous blood flow and hence insulin absorption [89]). The same may be true of physical activity, as it results in increased SC blood flow (to get rid of heat) and induces motion due to the vigorous use of the underlying muscles. The accumulation of SC fibrous tissue after years or decades of insulin injections may markedly increase insulin absorption variability. There is also interindividual variation, which may increase the need for individually adjusted algorithms in APs.

When using CSII, there is also the challenge of transporting the insulin from the ampulla in the device to the SC injection site. This is prone to several errors, which could be down to the user or due to equipment failure [90].

As with all patients, physical activity increases insulin sensitivity in patients with DM1 [91]. This is frequently observed in DM1 patients who regularly exercise in the evening. Hypoglycemic episodes often occur in such patients later in the evening or during the night. However, “unexplained” hypoglycemic episodes are also often experienced the next morning or later the next day. This may even happen in cases where the patient has increased food intake and/or reduced insulin doses due to the exercise.

There is no official recommendation regarding the amount or type of carbohydrate patients with DM1 should ingest to avoid hypoglycemia during exercise. A clamp study found IV glucose demand changed as the level of physical exercise was increased consecutively. The glucose infusion rate was increased at a moderately intense level of exercise, while the need for IV glucose disappeared during high-intensity exercise [92]. The latter illustrates the complexity involved in attempting to take physical activity into account, as the need for insulin can vary during exercise in ways that are difficult to predict.

It was recently reported that the effects of physical exercise on blood glucose depend on the nature of the physical exercise; i.e., its duration, degree of vigorousness, and the type of physical exercise performed [93]. It seems that the glucose level initially drops more quickly during aerobic activity than resistance activity, but resistance activity leads to a longer-lasting lowering of the glucose after the activity is finished [94]. Research groups who have accounted for physical activity in their algorithms have achieved lower hypoglycemic episode rates during physical activity, but long-term studies are needed to check that these lowered rates do not come at the cost of more time spent in hyperglycemia [95‐97].

Any AP with a feedback controller exclusively based on blood glucose levels would need to accept markedly less satisfactory glucose control in order to avoid both short- and medium-term hypoglycemic episodes due to physical activity. In particular, this problem would become more pressing if unpredicted major exercise was to be performed within the first few hours after a meal, as fast-acting “meal” insulin lowers glucose even 5 h after administration. Due to the relatively long half-life of SC-delivered insulin, this issue is difficult to resolve unless a compensating factor such as glucagon is applied [22, 29, 98].

Most studies of AP have used meal announcement, which involves the patient either deciding the amount of insulin given with a meal or estimating the amount of carbohydrates and leaving it to the algorithm of the AP to suggest the insulin to be given with the meal [21, 23‐35]. Meal-related insulin administration is a major challenge for both an AP and patients with DM1.

Only two smaller studies by a Dutch group have tested a bihormonal (insulin and glucagon) AP without meal announcements and user-dependent insulin determination [22, 98]. In the first study, an AP was used for only two days in a limited number of highly surveyed patients [22]. Interestingly, mean glucose levels did not improve compared to the those obtained with the usual open-loop treatment, while night-time glucose levels improved.

In healthy nondiabetic individuals, the first phase of insulin release (see the “Insulin Physiology” section) is probably important for restricting the total meal-induced increase in glucose. The larger release of insulin seen after about 10 min (the second phase) is essential for handling the increase in glucose within a couple of hours. Thus, it is essential for an AP to be able to identify a meal as soon as possible after the meal has been initiated if it is to deliver insulin early enough to achieve proper control of meal-induced glucose excursions. The time until automated meal detection is significantly influenced by the latency of the SC CGM if the detection algorithm is based on these commonly available measurements.

Several studies of double SC AP have used two SC glucose sensors, where the second sensor signal is not actively used until the first sensor fails [23, 26‐28]. One study did apply a rule where an alarm was triggered if the difference between the two sensors exceeded >20% when the glucose level was lower than 8 mmol/L, or if the difference between the sensors exceeded >1.5 mmol/L when the glucose level was above 8 mmol/L [22].

Some control methods follow certain safety rules; for instance, only a partial (pre-meal) bolus is given at the start of the meal in order to avoid postprandial hypoglycemia, and an extra post-meal bolus is delivered if hyperglycemia becomes imminent [26, 28]. Other efforts to avoid hypoglycemia allow the algorithm to adapt its insulin sensitivity factor based on the rate of change in postprandial glucose excursions, and/or to use a “withhold action threshold” by defining an upper limit for the rate of change in insulin delivery as compared to baseline insulin delivery [22, 28]. Some control methods also try to achieve safe insulin delivery at decreasing glucose values by, for instance, implementing a time-limited suspend at a defined hypoglycemic limit when hypoglycemia is present or imminent, and/or by simply alerting the user to the need to ingest carbohydrates at low glucose values [25, 26]. Some groups apply glucagon, typically by giving increasing pulses of glucagon at certain glucose “action thresholds” defined by glucose levels or defined changes in the rate of decrease in glucose [22, 29]. Other control methods also account for the potential “stacking” of previously delivered insulin [25‐27, 29].