Dosing Algorithms for Insulin Pumps

Abstract

Advances in insulin pumps and dosing algorithms have improved the regulation of blood glucose concentrations, reduced the frequency of hyperglycemic and hypoglycemic episodes, and improved quality of life for people with diabetes and their families. Whereas nonautomated or standard insulin pumps may work with a smartphone app to provide information to reduce the workload of users, major advances have been made in automated insulin delivery (AID) system pumps that adjust insulin infusion rates via proprietary algorithms in real time based on data from a continuous glucose monitoring sensor that communicates with the AID pump to take over additional tasks. Improvements in these algorithms are reducing the burden on users today, while the next generation of AID systems, with full automation, are under development.

Advances in various technologies during the past two decades have enabled the development of various types of insulin pumps and automated insulin delivery (AID) systems, sometimes also called artificial pancreas systems. Today, individuals with diabetes have several options ranging from insulin pumps operated manually to pumps with embedded AID systems and pumps that are controlled by apps operating on smartphones. Various smartphone apps provide information to reduce the number of tasks required of manual pump users. Additionally, evolving continuous glucose monitoring (CGM) sensor technology and advanced control algorithms have improved the performance of AID systems compared with sensor-augmented manual pumps.

Manual pump users can make their dosing decisions by receiving information and guidance from various smartphone apps that compute the carbohydrate content of their meals and suggest an appropriate insulin bolus dose. These apps can also be used by all individuals with diabetes regardless of their method of insulin delivery. Apps, including ones that can estimate carbohydrate amounts from pictures of the meals, nudge people to perform physical activity, and warn of glycemic trends toward hypoglycemia or hyperglycemia based on CGM data, have grown in popularity among people with diabetes because they reduce the self-management workload and the probability of having hypoglycemic or hyperglycemic episodes. Individuals who have a good understanding (i.e., a mental model) of how their body behaves in response to various factors that affect their glucose concentration (GC) can use that knowledge to adjust the insulin infusion rates of their pumps. This ability to personalize dosing decisions was an advantage of manual dosing decisions over some early-generation AID systems that took over limited automated tasks such as real-time basal adjustments.

Progress in refining AID system features and improving performance has accelerated significantly during the past decade. The promise of AID was demonstrated in clinical experiments since the 1970s (1), in which blood glucose values were generated from blood collected from a vein, an appropriate insulin dose was computed by a computer, and insulin was injected directly into a vein. Because direct access to veins was not acceptable for devices to be used in daily living, the achievement of viable AID systems took several decades, as methods of measuring GCs and infusing insulin into subcutaneous tissue were refined. Advances in CGM from subcutaneous tissue and the design and production of small, accurate insulin pumps complemented the development of novel control algorithms. Progress in research and development in both industry and academic institutions culminated in the commercial availability of the first hybrid closed-loop (HCL) AID system in 2018 (2). Large clinical experiments (3–6) and growing data from users of various AID systems available to people with type 1 diabetes (7–10) demonstrated that the glycemic time in range (TIR; 70–180 mg/dL) increased significantly with AID use compared with glucose regulation by multiple daily injection insulin therapy or the use of sensor-augmented pumps (SAPs), in which insulin delivery is not automated and all dosing decisions are made by users.

It took more than four decades to move AID technology from the research setting in a hospital room to daily living at home (11–15) and less than a decade for evolving HCL AID systems to document the achievement of 60–75% TIR. Yet, several challenges and opportunities still exist to increase TIR to ≥90% with AID. The challenges range from accounting for inter-individual differences in the dynamics of glucose metabolism, intra-individual variations in glucose metabolism from day to day, variability of insulin pharmacokinetics, and physical activity to capturing and mitigating the effects of special circumstances such as puberty, sickness, menstrual cycles, perimenopause, and unexpected, spontaneous events. Research efforts focused on addressing these challenges include 1) improvement of models predicting future glucose values and personalization of algorithms for each user, 2) incorporation of information from additional sensors to capture individuals’ metabolic activities that affect GC dynamics, and 3) use of a second hormone (glucagon) to better mimic the ability of the human pancreas to regulate GCs. The path of development is traveling from HCL to fully automated AID systems by leveraging new developments, including the production of novel devices and the application of machine learning techniques using artificial intelligence (AI) technology. The remainder of this article focuses on these issues.

Improving the Ability to Predict Near-Future Changes in Glucose Concentration

A very high TIR would no doubt be achieved by an individual who has a good understanding of the dynamics of his or her glucose level variations, can estimate the carbohydrate content of meals accurately and enter them ∼30 minutes before the start of meals, has a set daily routine for various physical activities throughout the day and can adjust glucose goals and insulin infusion rates well before beginning exercise sessions, experiences minimal anxieties and stress, and has a well-set sleep routine. The challenge is to develop a fully automated AID system that can achieve a similar level of TIR performance for individuals who may eat meals at different times every day, consume different amounts and types of foods, have both a set exercise routine and spontaneous physical activities, face frequent events causing anxiety and stress (e.g., road rage in rush hour traffic, the loss of a close friend, financial challenges, a house fire, or winning the lottery), and experience irregularities in sleep pattern and duration.

Current HCL AID systems rely on CGM data to inform automated basal insulin delivery and correction dosing for hyperglycemia, with some degree of manual user input for meals and exercise events. Their dosing algorithms use recent CGM data, active insulin (insulin recently infused that is still available [insulin on board]), and user-entered meal and/or exercise information to calculate the optimal infusion rate at any given time. The AID framework has the potential to automate information collection from additional sensors and to automate interpretation and decision-making (dosing algorithms) to maintain euglycemia despite myriad daily disturbances such as meals, physical activity, sleep, and stress.

The first step in shifting from HCL AID to full automation in insulin dosing is to leverage information from other sensors that individuals accept wearing in their regular lives. Today, many people wear wristbands (16–19) or rings (20) that can track their heart rates and physical movements in three dimensions (i.e., three-axis accelerometers). Some devices contain additional sensors to report skin temperature and galvanic skin response (indicating perspiration). These variables provide complementary information to current glycemic data about the occurrence and characteristics of disturbances that will affect glucose levels in the near future. For example, an increase in heart rate may indicate physical activity or anxiety. Because these two source causes affect glucose levels in opposite directions, it is important to determine the actual cause. Glucose concentrations are also affected by the duration and intensity of these events. Furthermore, sometimes multiple events may overlap in time, and the presence of all these events (i.e., disturbances) must be detected.

Although wristbands and rings provide clear reports of heart rate or minutes of physical activity, their sensors initially generate noisy signals affected by artifacts and may have missing values and outliers. Although a human can immediately notice a jump in CGM glucose value from 110 to 400 mg/dL in 5 minutes and discard the latter value as an outlier, such assessments must be programmed into an algorithm. Hence, the sensors and AIDs include software for data preprocessing, including noise filtration, artifact elimination, and imputation of missing values.

For AID systems, the data from these devices should be streaming in real time (as opposed to providing a summary at the end of a period or activity). For example, the AID system needs to detect the start of spontaneous physical activity within its first minute to initiate mitigation. These devices collect data in real time that are stored in the clouds of the device manufacturers, but the software that would enable individual users or an AID system to receive these data in real time is not usually provided to users.

Decision-making based on the values reported for a specific variable may not be reliable (recall the example of heart rate increase with exercise or stress). It is also practically impossible to detect concurrent disturbances by interpreting raw signal data. Features are generated from these signals, and the most informative features for specific disturbances, or those that can discriminate between different disturbances, are selected for use in disturbance detection, discrimination, and quantification. Various algorithms based on statistical techniques and machine learning are used for generating these features and then using the final set of features to detect and estimate the characteristics of various disturbances. Often, some derived variables can summarize the characteristics of a disturbance. For example, energy expenditure can be computed from these sensor data every minute and expressed as metabolic equivalent of task (MET) values. MET values indicate the intensity of the activity, which can be used to detect the presence of the exercise and determine its potential impact on glucose levels, enabling instantaneous decisions (e.g., stop an insulin bolus, reduce the basal rate, or shift the target glucose level automatically).

Factoring in Physical Activity

Physical activity includes structured exercise, activities of daily living, and spontaneous (unplanned) activities. Although users can enter structured exercise to inform an AID, activities or daily living and spontaneous activities are considered to be low intensity, split into small active periods during the day, or unpredictable. The intensity of any type of activity and its duration influence its effect. Low- and medium-intensity activities reduce the GC, whereas high-intensity and anaerobic activities can increase it. MET values provide a reliable indicator of intensity and duration of activity. An additional factor to consider is whether the physical activity is aerobic exercise or resistance exercise, as the former causes larger drop in GC than the latter. The effect of interval training falls between those of aerobic and resistance exercise. MET patterns and accelerometer data can discriminate these types of physical activity.

Detecting Acute Psychological Stress

Taking an exam, participating in a competition, sitting for a job interview, and speaking publicly are examples of anxiety-causing events that can affect the GC. Although chronic stress, which is sustained over long periods of time, has no impact on insulin dosing decisions made for the next 5 minutes or 2 hours, acute psychological stress affects the GC in this short time window (21). Features based on heart rate, accelerometer, and galvanic skin response data (22) can detect the presence of acute psychological stress, discriminate the type of stress (e.g., anxiety or mental stress), and determine whether the stress is occurring concurrently with another factor such as physical activity. A typical example would be anxiety before the beginning of a race and competitive stress during the race, with neither being present if the same type and intensity of physical activity is done in practice (21). Consequently, algorithms that detect, discriminate, and quantify physical activity and acute psychological stress will be crucial in the development of fully automated AID systems.

Assessing Variations in Sleep Patterns

Many individuals with diabetes have reported poor sleep such as short sleep duration, frequent sleep interruptions, hypoglycemia during sleep, and anxiety that affects the quality of sleep. Sleep characteristics affect insulin sensitivity during the morning hours of the next day. Sleep monitoring by various devices in the clinic or at home have been done for several decades. Several smartphone apps are available to monitor and assess sleep, with accuracies appropriate for use in insulin dosing decisions. The signals are derived from accelerometer and heart rate data, and the sleep information can be used in an AID system to compute insulin sensitivity the morning after sleep.

Detecting Meals and Estimating Carbohydrate Content

An insulin dosing algorithm must recognize the start of a meal and its carbohydrate content. Some recent HCL AID systems reduced the burden on users by limiting the type of information they need to input to just the meal start or estimated start time and a simpler meal content report, such as noting a small, medium, or large meal or specifying whether a meal is similar to the same meal on the previous day.

Various approaches have been proposed to automatically specify the start of a meal and its carbohydrate content. Early work focused on determining mealtimes and characteristics from historical data to predict the time and dose of insulin boluses (23,24). However, precautions must be included if the current eating behavior is different from the patterns captured. An extension of this approach is to capture the various eating patterns of an individual and use those patterns to select the individual’s current eating pattern. This information is then used to provide a sequence of mini-boluses of insulin as evidence builds over time on the usual occurrence of a meal and its carbohydrate content (25,26). The first mini-boluses are conservative, and if further evidence is not confirmed over time, bolusing is stopped. The aggressiveness of insulin mini-boluses increases as successive CGM values confirm a meal. Estimating current and future meals from historical and current data is further extended by leveraging AI to develop Bayesian networks, a type of probabilistic graphic model, and adjusting insulin infusions accordingly (27). A different approach relies on detecting the occurrence of a meal from current CGM data, estimating the meal’s carbohydrate content from recent CGM data, and suggesting insulin mini-boluses based on dynamic estimation of meal characteristics (28,29). Leveraging smartwatch accelerometer data, eating and drinking can be detected based on hand gestures in real time (30,31). It is realistic to expect that several of these techniques will be included in a dosing algorithm, and a consensus decision will be made for initiating and adjusting bolus insulin dosing accordingly.

Models for Predicting Future GC and Personalizing AID

All insulin dosing algorithms rely on mathematical models that describe GC dynamics in response to various factors. The models are a set of differential equations that quantify the effects of insulin, glucose, and other factors (e.g., physical activity and acute psychological stress) and variations in insulin sensitivity on GC dynamics. Most models consider the effects of meals on GC, and some recent models also include the effects of physical activity (21,32).

Many earlier models had a fixed set of parameters. As the limitations in achieving higher TIR were considered when using the same model for people having different metabolic characteristics, adaptation of the models was considered. Some AID algorithms adapt their model based on the previous day’s data. Others provide more frequent adaptation to capture the impact of a meal or physical activity on the parameters of the model. Recursive identification enables the development of adaptive models that are updated as frequently as the speed of CGM data reception (every 5 minutes) to maximize the model’s accuracy for GC prediction horizons (ranging from 30 minutes to 2 hours) used in most AID control algorithms (27,33–35).

Most AID systems use model predictive control (MPC) algorithms. The models in these MPC algorithms predict future GC values based on suggestions of future hypothetical insulin doses to find the optimal insulin adjustments. These GC predictions will be more accurate if the effects of future disturbances (meals, physical activities) are included in predicting future GC values. As discussed in detail below, an MPC algorithm uses estimates of the output (GC, in this case) in the future for specific (hypothetical) changes in future inputs (insulin doses infused) (35). Next, these future inputs are modified to seek future GC trajectories that will be as close as possible to the desired GC trajectory. When the best match of the estimated future GC trajectory to the reference trajectory is determined, the insulin dosing sequence that achieved it is selected, and the insulin pump is ordered to infuse its first value in the sequence (the dose to be infused in the next 5 minutes). Next, the whole computation and optimization process is repeated to determine the best insulin dose at the next infusion time. Because a new disturbance can affect the accuracy of these projections in future times, it is important to include future disturbances that may occur during the future time window of interest.

One way to incorporate the effects of future disturbances in control decisions is to capture the different patterns of behavior of an individual with diabetes from historical data (36,37). People have a finite number of behavior patterns (e.g., weekday, weekend, travel, vacation, and during sickness) that can be extracted from their historical datasets. Meal and exercise times and characteristics are used to describe these patterns. Then, on the current day, by matching the current day’s GC profile to one of those patterns, an educated guess is made about meals, physical activity, and their characteristics for the rest of the day and used in the estimations made about the future GC trajectories of the day.

Control Algorithms for Insulin Dosing

The pancreas of an individual without type 1 diabetes rapidly detects changes in GC and other hormones to make insulin secretion decisions. Insulin is released into the bloodstream near the liver and rapidly transported to all parts of the body to enable glucose absorption to various tissues. CGM-only AID systems react to increases in CGM values. They can learn from CGM data that the glucose level is changing, but they lack the additional information that is available to the pancreas, generated by various hormones and signals in the body about a meal to be consumed, or an exercise that has started, until the effects of those factors appear in the CGM data. Hence, the HCL AID framework was introduced to manually provide information to the AID system to let the controller start making adjustments for meals and exercise well before changes in CG values can be detected.

The insulin infusion to subcutaneous tissue by the AID system becomes effective within 40–50 minutes because of slow diffusion of insulin from subcutaneous tissue to the bloodstream. Hence, TIR achievable with a fully automated AID system would increase if the information on meals, physical activity, and other disturbances is made available as soon as they occur, well before the GC values start changing.

Multivariable AID Approach

The multivariable AID (mvAID) approach includes proactive decisions based on real-time information from wearable devices that reduce delays in the system learning about the presence of various disturbances and consequently accelerating decision-making to adjust insulin dosing to mitigate the effects of disturbances sooner (38). The current mvAID structure shown in Figure 1 illustrates the various components of the dosing algorithm. CGM data and data from a wristband are streamed in real time to a smartphone app. Two neural networks estimate the energy expenditure and detect the presence and determine the type of acute psychological stress. Meal information is generated by an unscented Kalman filter, which has a detailed metabolic model to interpret CGM data and detect the start time of a meal and its carbohydrate content. All these inputs generated by the modules of the mvAID system are fed to its MPC algorithm, which computes the future optimal insulin infusion rates. The infusion rate for the next time instance is sent to the pump to infuse the insulin dose computed. The sequence of these activities is repeated when the new CGM data are reported 5 minutes later. The wristband data are also used to detect the presence of physical activity every 1 minute and, upon detection, the system stops any insulin bolus, reduces basal insulin infusion, and modifies the glucose goal. The metabolic model in the MPC algorithm is an extended version of Hovorka’s model (39), with physical activity and acute psychological stress submodels.

Figure 1.

mvAID system. Data collection, data interpretation, estimation of future glucose trajectory, computation of the optimal insulin dose, and instruction to the pump for its delivery are shown. exer, exercise; g, grams; ins, insulin, str, stress, U, units.

The mvAID framework provides an insulin dosing algorithm that can accommodate unexpected events (because the sensors will provide appropriate information in real time) and enable a fully automated AID system with no manual input requirements. The MPC model is recursively updated to represent the current dynamics of glucose metabolism. It has guardrails to prevent very aggressive or slow responses.

Pattern-Based Approach

Another fully automated AID approach relies only on CGM data supplemented by users’ various daily living patterns, which capture their meal consumption habits and exercise routines. The current day’s CGM profile is matched to the glucose pattern and, as the CGM values start hinting at a meal or exercise, the algorithm makes a conservative adjustment: a mini-bolus in the case of a meal or a reduction in insulin infusion and adjustment of glucose goals for exercise. As evidence builds with new CDM data at subsequent sampling times, the interventions become more aggressive. If the next CGM values do not indicate the presence of the disturbance, then the changes in insulin infusion are stopped. Although the first few interventions may modify the GC, they do not cause hypoglycemia or hyperglycemia because the magnitude of adjustments is small. This approach requires a rich historical dataset and implicitly assumes no seasonal changes. Because most computations for pattern extraction from historical datasets are done offline, the computational load of this MPC algorithm is not too high. If users modify their lifestyle, the profiles need to be generated again. Treatment of unexpected events not captured in the patterns may be suboptimal.

Extensions of Insulin Dosing Algorithms

Do-It-Yourself Artificial Pancreas

The do-it-yourself (DIY) artificial pancreas system (APS) communities provided creative and user-friendly AID systems with which dosing decisions are made on a rule-based decision-making framework. DIY AID users have active online discussion groups and share tips and good practices. The versatile algorithms and user interfaces can be modified rapidly to fit the expectations of individual users. Several variants for these systems are available, including OpenAPS, AndroidAPS, and Loop (for iPhones) (40–42). These systems use CGM data and manually entered meal and physical activity information to generate insulin infusion values transmitted to an insulin pump. Tidepool has developed the Tidepool Loop algorithm, a version that is cleared by the U.S. Food and Drug Administration (FDA) and is available in the Sequel twiist AID system (43).

Dual-Hormone Artificial Pancreas Systems

Dual-hormone artificial pancreas systems provide a more accurate approximation of the functionality of the biological pancreas by administering both insulin and glucagon in response to changes in GC. Glucagon is given as mini-boluses to prevent or treat hypoglycemia while insulin infusion is suspended (44–48). One company has reported ongoing clinical trials for submission of results to the FDA (49).

Other Extensions

Other extensions of AID systems to accommodate pregnancy, gestational diabetes, the menstrual cycle, and perimenopause and coordination of AID operation with adjunctive therapies such as noninsulin medications for type 2 diabetes are areas of development now receiving attention (50–53).

Conclusion and Future Developments

The development of fully automated AID systems is progressing rapidly. The technology is feasible to develop algorithms that can identify and account for meals and physical activity, and the incorporation of mitigation of acute psychological stress is also within reach. One of the bottlenecks on the path to fully automated systems is the needed collaboration between diabetes technology companies and manufacturers of wristbands and rings for real-time streaming of data, both in research settings and for commercial production. The development of faster-acting insulins will also contribute in increasing the TIR achievable with AID. Dual-hormone and mvAID systems offer additional capabilities that will improve GC regulation and enable the attainment of stricter glycemic goals such as >90% TIR and further enhance the quality of life of individuals with diabetes.

The success of current AID systems in the management of type 1 diabetes has also enabled their consideration for use in type 2 diabetes (54), and the FDA has recently cleared one system for such use (55). The adjunctive use of various diabetes drugs such as GLP-1 and dual GLP-1/GIP receptor agonists with AID systems is another promising research direction to increase attainable TIR targets for people with diabetes.