Meeting Report: The Fourth Artificial Pancreas Workshop: Testing and Adoption of Current and Emerging Technologies

Testing of a Hybrid System and Transition to a Fully Automated System, Academic Perspective. Boris Kovatchev

In a change from the published agenda, this session was started off by Dr Boris Kovatchev, who described the contributions of a program, implemented in 2008 with FDA agreement, that uses a computer simulator of the human metabolic system (effectively replacing animal studies) and its impact on the AP field. This system, called in silico testing has been extremely valuable. A series of clinical trials of an artificial pancreas followed and in 2011 testing of a portable system using a smartphone as a controller was successful.

The UVA DIAS system is insulin only and modular, allowing processes running on different timescales to work simultaneously. Their modules control: (1) Meals and Exercise, (2) blood glucose (BG) corrections, (3) Basal and Steady state, and (4) Safety factors. Importantly, a hybrid system, needing input from the patient on food and exercise, differs from a fully automated system only in the meal/exercise module. Clinical studies of the hybrid systems have included younger children, as young as 5 years, and have become increasingly aggressive, with more patients, longer times, less oversight and more vigorous exercise. The JDRF CTR3 program was a study of continuous hybrid control for 6 months. This study was done in patients who were already in good glycemic control showed a reduction of HbA1c from 7.2% to 7.0% with a dramatic reduction in hypoglycemia and 75-85% of values were in the Time in Range (TIR) zone (between 70 mg/dl and 180 mg/dl). Ongoing studies include a study of overnight control alone, a study of children at a 5-day ski camp, and the iDCL definitive study, an NIH funded randomized study at 10 sites that includes 240 patients for 6 months of hybrid closed loop with sensor augmented pump (SAP) treatment as the control.

Transitioning to a fully automated system has challenges: (1) meal challenge due to the slow and persistent action of insulin, (2) exercise challenge due to sensor delays when detecting physical activity by BG drop and the inability of the pump and CGM signals to indicate changes in insulin sensitivity with exercise, and (3) intraday and interday variations in physiological parameters.

The next feature will be to develop a cloud based system that can use “big data” features to provide better control

Therapy Automation Evolution: Industry Perspective. Benyamin Grosman

Insulin pumps and continuous glucose monitoring (CGM) sensors lower HbA1c, each, by 0.5-0.75%. The Medtronic plan for an artificial pancreas is: (1) Threshold suspend, (2) Predictive threshold suspend, (3) Hybrid closed loop, (4) Advanced hybrid closed loop, and (5) Personalized closed loop. At the present time, the threshold suspend feature is available and working well. Predictive suspend has worked well in trials outside the United States. After testing a Blackberry-based basal control system, Medtronic has settled on an Android-based hybrid closed-loop system using a modified PID controller. Insulin infusion safeguards were designed around time limits, safe basal, missed communication, and sensor inaccuracy. In testing, the overnight control was very good. An unannounced meal, however, drove the blood glucose up to over 300 mg/dl. Unannounced exercise also resulted in more hypoglycemia, while missed sensor readings resulted in hyperglycemia.

Medtronic® has taken their hybrid closed loop through an uncontrolled pivotal trial including 124 subject and over 12 000 patient days. There was a reduction in HbA1c of 0.5% with no major adverse events. Although this seems like a small difference, the patients were very compliant and already on sensor augmented pumps and in good control before starting the hybrid closed loop. This device—the Medtronic MiniMed 670G—is now FDA approved.

In silico testing has also shown the advantage of an automated correction bolus added to a hybrid closed-loop system. Medtronic wants to personalize the system by adding context information, automating meal bolus completion and correction bolus to the system and incorporating IBM Watson®. Medtronic wants to take a holistic approach to diabetes management.

Single Hormone Systems. Roman Hovorka

Dr Roman Hovorka spoke about an insulin only artificial pancreas system: Single Hormone Systems. Insulin only experimental systems have been available since 2000 and have been tested in home since 2012. There was also a pivotal study of a commercial system (Medtronic) in 2015.

An insulin only system has low biological risk given the vast amount of information already known from insulin only systems. They are also less complex, using well-known insulin pumps and continuous sensors. Furthermore, they can be innovative without increasing the complexity of the device.

There are a series of randomized clinical trials that have shown the effectiveness of the single hormone systems compared to sensor augmented pumps. The Moshe Phillip group demonstrated that overnight, both children at a diabetes camp (n = 56) and adults at home (n = 24), had better glucose control with more time in range and fewer hypoglycemic episodes with an insulin only artificial pancreas. Using the Medtronic artificial pancreas in camp studies with about 20 patients, there was an improvement in the overnight control, like the Phillip studies, but little overall change when the system was used all day for 6 days. The UVA group also showed improved overnight control in 10 adults studied at a hotel; but, in children aged 5-9, the 24-hour glucose control was worse. However, in Dr Hovorka’s 12-week unsupervised trial of overnight control in children, aged 6-18 (n = 25), there was a 25% increase in time in range with no change in hypoglycemia. The same study in adults (n = 33) showed an 11% improvement in time in target with a small decrease in the area under the curve for hypoglycemia. Overall, these studies seem to show the advantages of an artificial pancreas, but the variation in the studies shows that the algorithm and site of study matter and that other patient factors may matter as well.

A number of organizations are trying to commercialize an artificial pancreas, using either insulin (single hormone) or insulin plus glucagon (dual hormone), shown in Appendix B.

The future may bring some improvements without increasing complexity, such as simplified meal dosing (for example using only a few categories of meals—small meal, medium meal, large meal), faster-acting insulins, insulin catheters that last longer and/or include the sensor, factory calibrated CGM, and more advanced algorithms. Other improvements may increase complexity, such as a second hormone and objects to “announce” exercise or improve exercise control, such as accelerometers or a heart rate monitor.

Large, well-controlled studies are needed and planned. There are large outcome studies planned or underway in the United States, Australia, and Europe with up to 200 patients for up to 2 years, using prototype systems. The NIH is funding the International Diabetes Closed Loop (iDCL) Trial in 10 centers in the United States and Europe with 240 patients studied over 6 months using the UVA/Type Zero modular algorithm and evaluating HbA1c and hypoglycemia. They will use patients that are in poorer control (HbA1c > 8.5%) and SAP therapy as the comparator. Finally, the Cambridge group will be studying their algorithm in a Medtronic system in 65 children and adolescents and 65 controls. This will be a 12 month day and night automated closed-loop insulin delivery, multicenter study (6) under free living conditions, and used standard of care as the comparator.

Dual/Multihormone Systems. Steven Russell

Dr Steven Russell discussed dual hormone systems. As he described, insulin alone AP systems are limited by the poor pharmacokinetics of currently available insulin formulations, the need for meal and exercise signals by the patient, and postprandial early hyper- and late hypoglycemia. Patients with type 1 diabetes are also deficient in both amylin and glucagon; as such, adding a second hormone theoretically might help glycemic control. Amylin, GLP-1 agonists, and SGLT inhibitors limit postprandial hyperglycemia, but glucagon reduces hypoglycemia and allows more aggressive insulin dosing to achieve lower glycemic targets. A dual hormone system using glucagon makes the system more complex to design, but may make it easier to use and safer.

Pramlintide and GLP-1 have been studied as second hormones in a closed loop by the Yale group. Pramlintide lowered the glucose values somewhat and delayed the meal peaks. Liraglutide (GLP-1 agonist) had a modest effect of lowering blood glucose but reduced the insulin dosing by 25%.

Glucagon is being added as a second hormone by four groups: McGill, Boston U/MGH/Beta Bionics, AMC/Inreda Diabetic, and Oregon OHSU. The McGill system is a simple addition of glucagon to reduce hypoglycemia on a system designed primarily for insulin. The BU system is a fully automated system (requiring no meal data) or minimal meal data and is designed as a bihormonal system. The AMC system is fully automated and also designed as a bihormonal system.

The McGill group showed that in an insulin only overnight study, their version of the artificial pancreas substantially lowered glucose and increased time in range. The addition of glucagon improved this slightly (TIR: usual care 54%, insulin only 77%, bihormonal 84%). They have a large number of studies ongoing and planned.

Using the Boston system, an insulin-only and a bihormonal system perform similarly at a target BG of 130 (time in range: usual control 62%, insulin only 67%, bihormonal 71%). However, with lower glucose target ranges, down to 100 mg/dL, the time in range improves substantially using the bihormonal system (bihormonal 81%) without increasing the risk of hypoglycemia. (Note: with the lower glycemic targets, the infused glucagon rates can increase to 8.3 mg/kg/day.) When compared on how well they achieve the DCCT experimental group mean BG of slightly more than 150 mg/dL, usual care and insulin only achieve this 5% of the time, whereas the dual system with a target of 100 achieves this 90% of the time.

Beta Bionics (insulin only) and Inreda plan to have regulatory submissions in 2017. At least three companies are working on a stable glucagon (Zealand, Xeris, and Adocia) and approval of a stable glucagon must precede approval of a dual hormone system.

Remaining questions include: (1) Safety of chronic micro-doses of glucagon; (2) defining appropriate populations for artificial pancreas and for dual hormone AP system testing; (3) hybrid versus fully automated systems; (4) methods of pramlintide delivery; and (5) side effects of multihormonal approaches.

Summary of Main Discussion Points

The panel consisted of Stacey Anderson, Ananda Basu, Roy Beck. Claudio Cobelli, J.H. de Vries, Ahmad Haidar, Roman Hovorka, Moshe Phillip, John Pickup, Steven Russell and William Tamborlane.

Why Are Psychosocial Issues Crucial? Use of New Measures. Katharine Barnard

The success of an artificial pancreas is dependent on the patient’s perception that the device is useful and their willingness to use it. Dr Barnard and her colleagues are developing measures that assess the psychosocial impact of automated insulin delivery devices. These measures should be useful in both regulatory and reimbursement decisions.

Most patients take safety and effectiveness for granted; they are concerned about ease of use, improving their life and not thinking about diabetes all of the time. “This is not about glycemic control. It’s about, ‘How can I use this to improve my life or to improve the burden of diabetes?’” Thus, patients with T1D focus on practicality and usability. Product should be good the first time because a bad reputation is hard to overcome. Patients still complain about disappointing characteristics of older CGM systems. Furthermore, the World Health Organization (WHO) defines health as the absence of disease and the presence of physical, MENTAL and SOCIAL well-being.

It is important not to hype the artificial pancreas to levels not actually attainable as this will disappoint patients enough that they may not want to continue to use it. As investigators look to measures of practicality and usability, they can start with the FDA definition of a Patient Reported Outcome. These are raw reports directly from the patient that can be measured absolutely or as a change or in a clinical trial the measured effect on a concept (symptom, function, group of functions).

It is important to know whether the patient will accept the device, and what are the pros and cons of the device. Do they trust it and can they use it? What is the impact on social relationships, family, members and on the burden of the disease?

Across all clinical trials of the artificial pancreas systems, investigators should measure psychosocial outcomes and process evaluations at the same time points and with the same analytical approach. Furthermore, trials will be expected to report all results. Quality of Life (QOL) benefits derive from greater confidence in using the device, perceived control over the patient’s diabetes, satisfaction with the device performance, usability, acceptability and trusting the device. This may be a challenge with less technically skilled patients. Device design should include these factors which will increase use, overcome barriers, and increase confidence in the device.

Hyping the device can dissuade continued use. Investigators acknowledge and expect that the first AP systems will have limited features. But, if expectations are too high (eg, like they were for Exubera), then there may not be second-generation systems.

Psychosocial research will be part of the large artificial pancreas trial, measuring attitudes, beliefs, and emotions of the subjects. The INSPIRE measures have been developed, piloted and are undergoing reliability and validation. They should be widely available soon. Key concepts include concerns about the burdens, features, finances, medical benefit, context, interactions with the system, night time effects, QOL, technology, and trust. If successful, INSPIRE will allow critical appraisal of patient reported outcomes (as required by the FDA), will help form clinical practice around barriers, and will help payers understand the impact of the devices.

Potential Psychological Impact of a Long-Term AP System. Linda Gonder-Frederick

Since there are no long-term studies of AP use in clinical practice, the talk by Dr Gonder-Frederick was largely speculative in nature. She reported that the next step in AP clinical trials will be the collection of patient reported outcomes and qualitative data on patient experiences. Up to now, the studies have been done with patients under good control, who have little difficulty with their diabetes management. The next step in trials will be with the other patients, who have difficulty with control, lower literacy and less technical savvy.

Researchers now study aspects of blood glucose control and risk of complications. From a patient’s perspective those are important, but so is making diabetes management easier and improved QOL. Some studies of AP have shown a positive psychological impact: better QOL, less fear of hypoglycemia, reduced diabetes-related stress, sleeping better, and feeling better in the morning.

However, there are potentially significant barriers to long-term use of AP systems: Physical and emotional burden, negative response to wearing big obvious devices, achieving and maintaining a necessary skill set, burnout, and lack of reinforcement (perceived personal benefit). Furthermore, the diffusion of diabetes technology is slow and 50% of the population has an IQ less than 100. Therefore, the success of an AP system may well depend on keeping it simple. Most people do not go through a rational decision making process but rather a “tradeoff analysis” using a perceived benefit/cost analysis. People respond well to positive reinforcement. Thus, providers need to integrate positive reinforcement systems into the AP platforms to increase the benefit/cost ratio.

The medical community may also be able to learn from behavioral internet interventions. These start by having the patient set personally meaningful goals with measurable outcomes. Investigators then measure the outcomes, provide feedback and reinforcement and provide advisory feedback and support for problem solving to help patients reach their goal or modify it.

As providers use these systems more, they will see some challenging behavioral effects: people testing the limits of the system (ie, dietary indiscretion, exercise, partying), the phenomenon that some individuals “can’t unlearn” certain behaviors that are no longer necessary with the AP, conflicting information (ie, symptoms don’t match numbers), and finally too much focus on insulin with less on other diabetes management behaviors. To maximize psychological and clinical outcomes, providers need to give the patient appropriate education, clinical, and psychological support and intervene with problem solving therapies as needed. Collecting psychological data in the newer long-term trials will allow providers to better serve the patients.

Impact on Quality of Life: Need for Standard Measurements for Patients and Caretakers? Lori Laffel

Quality of life (QOL) is a key patient-reported outcome and an assessment of well-being. It consists of generic QOL, not related to health, and diabetes-specific QOL related to general health, diabetes management, and diabetes control.

Diabetes can place an enormous burden on patients and their families: insulin administration, carb counting, blood glucose measurements, worrying about exercise, diabetes education and communication with the diabetes team, school and family, each often many times per day, 24 hours per day, 7 days a week. This enormous burden can affect the QOL.

Type 1 diabetes (T1D) is not a do-it-yourself disease. Others are usually involved. It is stressful, with physical and emotion demands. There is guilt at diagnosis, a series of management tasks to learn, and constant measurement of adherence and evaluating new treatment tools. There are social worries of growth and development, fitting in at school, work, and socially, reproductive health concerns, and worries about long-term complications. There are mental and physical demands of multiple doses of insulin, BGM checks, and carb counting. Patients need to manage exercise, treat low BG and manage sick days. They have a lot of supplies to carry or wear and have frequent medical follow up. There is never a break, they feel different, and it is hard to be spontaneous. Patients feel frustrated, overwhelmed, and angry. They have a fear of both high and low BG, have to cope with the “Diabetes Police” and often have mood fluctuation, depression, and anxiety. This often leads to burnout. As investigators develop new AP technologies and platforms, there is a need for uniform practical measures.

Investigators currently have great measures of diabetes BG control. However, what is lacking is a QOL measure that performs consistently across people, place, and time. Components of a diabetes QOL include school or work, family, friends, mood, and treatment. It is applicable to the patient, parents or spouse, family, and the health team. QOL should also be followed over time.

Quality of Life domains include: social, relationships, family, work, community, emotional well-being, and health status. With newer, more complex devices, providers need to also include the burden of care and burnout. Burnout is common in chronic disease: the patient works hard without measurable benefit. This leads to feelings of helplessness, hopelessness, irritability and hostility, all leading to emotional exhaustion and physical or mental collapse. In T1D this comes for the unremitting nature of the disease, unrealistic goals for BG control with imperfect tools, leading to chronic frustration and failure and often repeated criticism from family, health care providers, and themselves. There is blame, worry, threat, and fear.

In the INSPIRE trial, there will be questionnaires to fully capture psychosocial aspects of using an artificial insulin delivery device. In the first stage of development there were 60 focus groups, with >400 participants and 89 individual interviews. Participants included patients with T1D of all ages, parents and partners. A coding group identified the main themes. Our group learned that patients have high expectations for an improved QOL on an artificial insulin delivery device. They expect the device to improve control with less thinking and less time, improve sleep, and improve their mood. They want less conflict and less worry and a reduced burden. However, they have concerns about accuracy, false alerts, device failures, multiple devices, frequent site changes, backup supplies and that they may need to take off the device for a shower or to go swimming. The device needs to be discreet, compact and unobtrusive.

Discussion

Toward an Adaptive Artificial Pancreas: From In Silico to Outpatient Free-Living. Claudio Cobelli

Adaptation is important as patients move from using the AP for a few days to wearing it for months. In a short trial, controller adaptation is superfluous but over time, interday and cyclical changes require adaptation. Our DIAS AP model uses a modular model predictive controller. The physical interface with the patient is the lowest module. The safety module is next with the real-time controller the highest layer in the standard system. In the adaptive system, there is a run-to-run adaptive module (R2R) above the real-time controller

The R2R module uses an individual’s recent past history to adjust the key variables (individual mealtime carb/insulin ratio and basal insulin delivery) to minimize time below 70 mg/dl (primary) and increase the time in target and move their mean glucose toward target (secondary). We have tested the R2R module in silico for two months; the system also considers day to day variability and the dawn phenomenon. Patients ate 3 meals per day, the carb/insulin ratio was randomly varied by 20% and insulin sensitivity by 30%. The R2R system improved control over the 2 months, lowering glucose by 10 mg/dl, improving time in range by 10%, and lowering time above 180 mg/dl.

We then tested the R2R system in free-living patients by extending a 4-month study with 4 weeks of nonadaptive controller followed by 4 weeks with a R2R controller. Moving from in silico to actual patients was very challenging. A computer model is very compliant, following all instructions; patients are less compliant, often doing what they want. The R2R module assumes no meals and no boluses overnight, but patients did both. The R2R module also assumes no additional food or insulin for 3 hours after a meal, but patients had snacks and did correction boluses. Patients also changed the controller suggestions, disconnected, had technical problems and just stopped the system. Thus, the rate of successful updates overnight was 62%. The remainder could not be updated because of manual intervention, insufficient data or meal presence. In contrast, the rate of successful updates during the day was only 23% with the same problems in the remaining data. Despite this, control improved overnight (TIR 64%>74%, P = .03, time above 180 34>24, P = .03) but remained the same during the day (moderate but not statistically significant improvements). If the same data are looked at over time, all of the factors improved significantly. What we learned is that we need to improve the algorithm to account for real life event and need to do longer testing.

Physiological Input Beyond Glucose: Relevance for More Personalized Systems. Marc Breton

To automate an AP response to exercise we need to gather information. We can use accelerometry, heart rate and galvanic skin response. Naïve AP systems are very good at preventing hypoglycemia except during and after exercise. This is because in an insulin only system, by the time the blood glucose is starting down after exercise, it is too late to intervene. Initially we used an uncomfortable harness that gave us accelerometry, heart rate, and respiration. For children playing soccer for 2 hours and using a low glucose suspend on detection of exercise, there was a slight effect at late that was lost by the end of the exercise. We used heart rate to switch our AP into an “exercise” mode for both adults and children. For a 30-minute exercise, in adults, standard AP fell about 30 mg/dl while the exercise-informed AP does not change. Similar results were found in children, but with more variability. Interestingly, an exercise module in DIAS improved time in range and overnight control. With an exercise mode in dual hormone system, there was a reduction in postexercise hypoglycemia (but not during exercise) but an increase in overnight hyperglycemia.

The group at the Illinois Institute of Technology has developed a system that uses heart rate, accelerometry and galvanic skin response. Their system takes into account the effect of exercise on insulin response as well as the changes in blood glucose. They have not done a comparison with and without the exercise system, but their data does show an appropriate reduction in insulin in response to exercise.

In order to study more vigorous exercise, we tested 32 adolescents at high-altitude in a ski camp. They were randomized to either AP or sensor augmented pump during 5 hours of skiing and other exercise. Overnight control was clearly superior with the AP but daytime control was similar. We were able to reduce hypoglycemia in the AP group. We manually let the AP know about exercise, but one of the subjects was wearing a Fitbit. Interestingly, hypoglycemia seems to lag exercise.

Finally, we have looked at some other factors. Sue Brown studied insulin sensitivity during the menstrual cycle and found significant differences during the various phases. This is being repeated and tested in silico.

Real-World Testing in Children and Adolescents. Bruce Buckingham

The vast majority of people with diabetes get it before age 15, so we can expect to start the AP in children and adolescents. There are lots of studies in children. In our low-glucose suspend study we started at age 3, Roman Hovorka has done AP studies in children as young as 6, and the Medtronic 670G studies included patient as young as 6 too. From the T1D Exchange data, we know that HbA1c reaches a peak of ~9% at age 15 and doesn’t return to a stable baseline of about 7.5% until age 30. Interestingly, even in adolescents, both insulin only and bihormonal AP systems produce an HbA1c of 6.5% to 7%. For myriad reasons, children have a high rate of severe hypoglycemia and DKA on pumps; moreover, for physiological reasons, the insulin requirement peaks in females at about age 12 and in males at about age 14.

A number of groups have found the accuracy of CGM systems in children to be much lower in clinical use than in the clinic. However, we noted an improvement when using a second drop of blood to calibrate the CGM system and it appears that this phenomenon is due to a lack of handwashing before calibration. The data plotted on an error grid shows that first drop calibration is typically skewed, but that the use of a second drop of blood for calibration was both accurate and unbiased.

A blunted adrenalin response to hypoglycemia is also common (30%) in children. Parents, in a blinded study, also failed to recognize hypoglycemia in their children 71% of the time. Remote monitoring can thus be very important.

Glycemic management in adolescents is further complicated by hormonal fluctuations, missed meals (but adolescents have huge appetites), missed insulin doses (65% of adolescents miss at least one dose per week), and skipped monitoring. Adolescents also have self-image problems and want to fit in.

Infants and toddlers have different challenges. They lack skin space for CGM systems and insulin pumps, have low insulin requirements, have irregular and often incomplete meals (plus there is no reliable way to quantitate breast feeding), and have frequent infections. Children, especially 4-10 year olds, also have frequent morning ketosis due to insufficient muscle mass to create enough alanine during an overnight fast. Furthermore, a prior study has shown that wearing a low glucose suspend system resulted in adverse skin reactions in 64% of children 3-6 years old.

Hypoglycemia presents a special challenge in infancy too. Infants cannot communicate their symptoms, have a smaller adrenaline response, are more likely to have seizures and hypoglycemia—both of which may have long-lasting adverse consequences on their developing brains.

Infants and toddlers also require very little insulin, with an average total daily dose of 4.5 units and a basal requirement of only 1.8 units. The basal rate averages 0.08 U/hr, and the average pump delivers insulin in packets of 0.025 U. Thus, a child may only be getting a bolus delivery every 20 minutes and the minimum change in bolus rates would be to 0.100 or 0.050 U/hr, a 33% change. We need better pumps.

Thus, closed-loop systems are appropriate for infants as they have the potential to: decrease hypoglycemia and hyperglycemia, improve sleep, and decrease the burden of diabetes. Adolescents are good candidates for closed-loop systems: they have the most to gain, since they are often in such poor control, and they usually have a parent at home to help them with the device. Inpatient clinical testing in toddlers and young children is hard for myriad reasons; rather, outpatient testing is optimal, and close-loop testing in this population should done in their home with appropriate backup for the parents.

Real-World Testing During Pregnancy. Helen Murphy

It is also important to include pregnant women in AP studies.

For myriad reasons, pregnant women have blood glucose control problems during pregnancy. On average, acceptable blood glucose control is attained in ~40% of individuals with type 1 diabetes initially and 50-55% at term. For individuals with type 2 diabetes, only ~55% achieve acceptable blood glucose control initially and ~70% at term. Glucose metabolism is different during pregnancy; the physiology is designed to ensure glucose delivery to the fetus. However, glucose disposal is limited in late gestation. This is advantageous for the baby but problematic for the mother with type 1 diabetes. Current insulins also create another problem. Insulin absorption slows and is more variable in late pregnancy, making glucose control harder.

There are still frequent complications of a diabetic pregnancy. Stillbirths or neonatal death occur in 6-10% of pregnancies, preterm delivery in almost 40%, neonatal ICU care in 40%, and overweight and obese infants in almost half. Most complications are associated with maternal hyperglycemia. Pregnant women spend about half of their time with glucose out of range. There is increasing use of CSII and CGMS and they have been proven to improve outcomes.

We tested the Hovorka algorithm in pregnant patients with carbohydrate amounts outside of the normal limits of pregnancy and with a half hour of exercise. Both in clinic and at home, the closed loop was effective in maintaining glucose in range (63-144 mg/dl). Importantly with respect to the generalizability of the results, most of the patients were both pump and CGM system naïve. We have tested this for a month at home. Interim results overnight showed: mean glucose 133>119 mg/dl (P < .01), time in range 60%>75% (P < .002), % time in hypoglycemia 1.9>1.4% (NS). The daytime results were: mean glucose 137>128 mg/dl (P < .0001), time in range 57%>66% (P < .0001), with no significant differences in hypoglycemia, total daily insulin dose, or hours of wearing the sensor. There was little difference in people who had experience with pumps compared to those on MDI. Burdens included: CGM system problems (47 events), pump problems (21 events), computer problems (13 events), and downloading problems (14 events). We trained the patient and the partners. It took 30-60 minutes to train them. At the time of delivery patients were in range 87% of the time with only 0.5% in hypoglycemic ranges and the mean glucose blood level was 109 mg/dl.

Real-World Testing in Older Adults. Richard Bergenstal

Only 20 people over 65 have been studied with a hybrid closed-loop system. The U.S. population is aging and baby-boomers are now above 65. As diabetes providers, we want and need to prevent complications and hypoglycemia. The T1D exchange shows only 29% of patients over 50 meet a target HbA1c of 7% but 20% have had a coma or seizure in the last 12 months. The incidence of seizure and coma increases with both age and duration in type 1 diabetes. The risk seems to be associated both with high glycemic variability and hypoglycemia unawareness, and perhaps beta blockers.

Type 1 diabetes in older individuals also seems to be associated with dementia. How will older patients handle the technology needed for an AP? How do you evaluate them? How to train them? Carb counting is hard.

We need standard measures of success. We use the ambulatory glucose profile, which is similar to the measures already being used. It is particularly good for longitudinal studies

Discussion

Introduction

Dr David Klonoff, Medical Director, Diabetes Research Institute, Mills-Peninsula Health Services and President of Diabetes Technology Society, welcomed the attendees for the second day. He briefly summarized the first day’s conclusions and described the second day’s agenda.

Regulatory Considerations for Component AP Systems

Discussion

Industrial Automation Interoperability Experiences: A Learning Opportunity for AP? Lane Desborough

Closed-Loop Data Management and Big Data Resources. Pratik Agrawal

Diabetes is both a burdensome and data driven disease. Patients often use their data to personalize their own therapy. To help patients with this, we at Medtronic have a roadmap outlined by Benyamin Grosman. Our job is to create the data pathways for a personalized closed loop. There are two components: (1) Collecting all of the non-pump data like food, exercise, sleep, stress, and biological cycles (eg, menstrual cycle data), and (2) Processing the data and making it actionable.

With threshold suspend alone, we were able to decrease serious hypoglycemic episodes by 5.3 per patient per year in over 18 000 patients. For better connectivity we plan to add BLE to newer systems, making it easier for patients to collect and report the pump data. With Carelink, we hope to collect other factors like sleep, stress, exercise, etc, but we do not yet have good data on their effects. More data should allow us to use big data techniques.

There are some concerns about big data. There can be: (1) Overload—how much computing do we need to do on a real-time basis, (2) Safety of the algorithms, (3) Security and integrity of the data, and (4) Privacy concerns. Overall, we need to go from data to information to knowledge to wisdom about how to use these systems.

In partnership with IBM Watson, we have tried to do hypoglycemia prediction. First we found that age, years with diabetes, and years on insulin were important variables. We were able to predict hypoglycemia (area under the curve by about 85% over 2 hours and 80% over 4 hours), and using this adaptive self-learning model, we have tested this in over 10 000 with 75-86% predictive accuracy.

Our model for the future is a series of “sensors” that send all of the needed data to a single location where patterns are recognized and processed by an algorithm that is continuously self-updating and sending instructions back to the pump.

Mobile Systems and Remote Monitoring and the Cloud-Connected Artificial Pancreas. Patrick Keith-Hynes

Cloud connected AP systems process a lot of data: (1) Patient data, such as risk monitoring, therapy optimization, education, and training, and (2) System data, such as system performance analysis and continuous improvement. Eventually, all of this will be driven by a cloud-connected treatment system.

In a cloud-connected AP system, the AP system must be capable of working on its own. The AP system collects data from the patient and uploads it to the cloud, where big data systems may be able to learn something. The cloud periodically feeds back to the patient, provider, payer, or the AP system for system improvements and the cycle repeats.

Patient data can: (1) Drive improvements in care with notification of pending glycemic directions and data archiving; (2) Foster patient-provider interactions through portals; (3) Optimize therapy, spot patterns and long-term trends, and (4) Integrate with additional data sources. System data can: (1) Drive continuous improvement of the algorithm; (2) Provide updates; (3) Deploy cloud-based enhancements; (4) Give patient feedback; and (5) Provide a pathway toward ubiquitous computing. Furthermore, there are several types of cloud-based systems: (1) Controller based, (2) Pump-based, or (3) Sensor based.

There are multiple benefits for cloud connectivity of an AP system. The immense processing power of the server and the big data analysis offers significant supplement to the stand-alone AP system. Patients may expect improvements in glycemic control and insights into their behaviors. Caregivers may expect reduced stress. Clinicians may expect improved glycemic control and better communication with patients. System providers may expect early warning of potential failures, gathering of data, and efficient software deployment. Finally, payers may expect the ability to measure the effectiveness of the AP system in providing improved diabetes care and insights into populations requiring special care.

Cloud connection allows multiple gateways into the data. Patients, caregivers, and providers can use the “User Gateway” to evaluate the data, especially at potential trouble zones, but can also interrogate “what if” scenarios using their own data to help give an answer. Caregivers have access to real-time data and to patient location if there is a problem. Providers may want to use the cloud to study the data from the system and to tweak it for better control. Payers may be interested in patient monitoring and data archiving, studying the effects of therapy optimization and following long-term trends and progression. There is also an AP gateway used by the system provider for continuous performance analysis, fault detection, and continuous software delivery. The AP gateway could also be broadened into a device gateway to include patients not on an AP but also electronic pens.

Cybersecurity Considerations for an Integrated System. David Klonoff

As we integrate diabetes technology with the internet, we encounter flaws in software that make our devices vulnerable to hacking. We need cybersecurity because security equals safety. Let’s start with some definitions: An Integrated Medical System is a device that monitors and transmits data and/or commands from or to a person connected with a hub. Cybersecurity is the protection of data and command information that are transmitted between connected medical devices. In diabetes therapy we have 4 devices that may be part of an integrated system: blood glucose monitors, continuous glucose monitors, insulin pumps (or electronic pens), and the artificial pancreas, which integrates the other 3.

There have been recent cybersecurity “milestones,” episodes in which our medical devices have been shown to be vulnerable to hacking. In 2008, a group reported hacking a pacemaker and cardiac defibrillator at the IEEE symposium on security. In 2011, insulin pumps were shown to be hackable.

The first attempts to regulate potential vulnerabilities occurred in 2014, with the FDA guidance for premarket submissions for management of cybersecurity in medical devices. These basic principles included: (1) Identify potential threats and protect against them; (2) Limit access to trusted users; (3) Ensure trusted content; (4) Detect intrusions, respond to them and recover function, and (5) Types of documentation of cybersecurity. To ease the regulatory burden, the FDA agreed to skip a review of software changes made solely to strengthen cybersecurity. In 2015, the FDA removed an infusion pump from the market because of poor cybersecurity. In 2016, they released a guidance for postmarket management of cybersecurity in medical devices. This document did not cover privacy risks, was not a safer harbor (would not alone protect the manufacturer from regulatory action or law suits), and stated that postmarket risk management is not a substitute for premarket risk assessment.

Hacking can take many forms, not all of them bad. As discussed by Howard Look, there are three examples of Do It Yourself hacking, where patients have hacked into their devices to gain access to their data or to take control of the device. Nightscout hacked into CGM systems to send the data to an easier to use reader. DIYPS is a couple in the state of Washington that hacked into systems to make an artificial pancreas for themselves. Their data are found in the third group, the Open APS system. This latter group has now had over 100 people create their own artificial pancreas systems. WebMD had a headline of “Hacking a Diabetes Cure?.” Of the first 40 users, 18 reported their data, with A1c 7.1→6.2%, time in range 58→81%, and sleep quality 94%.

The dark side is malicious hacking, most of which is hypothetical. In the IEEE spectrum, they point out that it would be possible. Ransomware is also possible. A hacker can take control of a device and require you to pay to have it released. Hollywood Presbyterian Hospital in California was hacked and had to pay ransom to have their data released. It has happened elsewhere: Kansas Heart Hospital and others in the United States, Canada, and Germany.

Last year, I went to the FDA and offered to have the Diabetes Technology Society build a diabetes cybersecurity standard. They agreed and in July of 2015 we held our first meeting with a broad consensus group of people involved in diabetes. We had: 6 government agencies—FDA, Homeland Security, NIS, NIH, NASA, and the USAF; four NGOs—ADA, JDRF, AADE, and the Endocrine Society; industry—diabetes companies—Agamatrix, Ascencia, Biorasis, Insulet, Sanofi, and Senseonics, and other companies—Blackberry, IBM, and Intel; and academics including diabetes experts, engineering, law, mathematics, and nursing. We also had patients and hackers, two cybersecurity firms, and an ex-CIA agent.

We used the “Common Criteria,” a standard for assessing the performance of software. It was developed by 6 countries (including the United States), is accepted by 26 countries, and has two basic requirements: Security and Assurance. All claims of security need to pass assurance testing to be certain they work. There are 6 levels of Assurance in the Common Criteria. We demand a high level, 4, since lives may be at stake. Because of the hard work of the committee, within 6 months (in December 2015) we released our standards for public comment and received many. We released our final standard in May 2016. Based on our recommendations, after a device is prepared according to our standard, it is taken to a certified third party assurance lab that will do vulnerability testing and penetration testing to see if they can hack into it. If it passes, the device is certified. Thus, there are security requirements, defined by a protection profile and the assurance requirements of laboratory testing.

I think we will have industry acceptance of the standard and more products will be tested for assurance of their security profile. This will be driven by a desire of stakeholders, including regulators for more security and bad publicity and lawsuits if security is too lax.

Discussion

Continuous Glucose Monitoring in Future Automated Insulin Delivery Systems—What Is Needed? W. Kenneth Ward

The characteristics of a perfect automated insulin delivery (AID) device are:

Sensor accuracy is important, but so is the delay. The delay is 5-6 minutes in normal and slightly longer in patients with type 1 diabetes. A thick sensor coating can add an addition 6-8 minutes and data processing still longer. The lag is less when glucose levels are falling, in part due to the increased insulin-induced glucose flux. In one study, we found the delay to be about 9 minutes when glucose levels were rising and 1.5 minutes when the glucose levels where falling.

All of the major methods of determining accuracy see to work: the error grids (Classical, Consensus, and Continuous), signed error values for bias, Bland Altman, and MARD. The best CGM systems currently are about 10% mean ARD. Some believe the best metric is the fraction of egregious errors (>30% or 30 mg/dl). Redundancy helps, but less so with new better sensors.

Our group has been very interested in measuring glucose levels at the site of insulin delivery, that is, with a single catheter. There are a series of papers that suggest it is possible. So, we wrapped an array of 6 sensors around a catheter. In pigs, with a low rate of insulin infusion, we saw little difference between a sensor wrapped around the insulin delivery catheter and a distant sensor. With large boluses of insulin, however, there was a substantial false peak in measured glucose concentrations in about 30% of the sensors. Insulin without preservative (phenol) did not cause this, but the preservative alone did.

There is a group in Europe doing the same thing, using an optical sensor (SPIDIMAN, Graz Austria).

Errors come from a few places. With an electrochemical sensor, the interfering substances remain: acetaminophen, ascorbic acid, and uric acid, all oxidized directly at the electrode. This can be mitigated by using a lower voltage device and/or different membranes. Pressure also reduces blood flow and gives false low readings.

There are some other novel CGM technologies. GlySense uses a glucose oxidase electrode, but measures oxygen utilization, rather than hydrogen peroxide production. They have tested fully implantable devices in people for over a year. Senseonics uses a boronate-based fluorescent electrode. In their clinical trial, 82% of the participants had the electrode implanted for at least 90 days with an HbA1c reduction from 7.6% to 7.1%. Their sensor has a mean ARD of 11.6%. The Agamatrix device does not require a needle for insertion of the sensor, and Profusa is interrogating optical interrogation of glucose oxidase.

What Is Needed for a High-Performance Miniaturized AGCS? Eyal Dassau

There are many facets that can be improved as we move to a high performance AGCS. The algorithm could be better, the insulin could be faster, the size of the system could be smaller, the placement of the system could be different (implantable?), automation of meal doses and exercise (user involvement) could improve, and device integration (sensor and pump) and data integration (cloud) could be better optimized.

Where are we now? There are multiple algorithms such as MPC, PID, and Fuzzy logic. All seem to function well. Some do better under stressful conditions, like an unannounced meal. We can get better.

Initialization should be easier. We demonstrated that by analyzing open loop data, we could reduce HbA1c faster. Medical devices shrink in half every 10 years. Pump sizes continue to shrink. SFC fluidics is working on a miniature pump that can deliver 2 hormones.

Controllers are being integrated into smartphones and embedded into the pumps themselves. Smartphone devices are powerful with large displays and built-in connectivity. However, patients may misplace them, battery life is short, they may be complex, and there are issues with updates, cybersecurity, and privacy. Embedded controllers have dedicated batteries and secure communications. They are an “on your body eco-system” and will not be lost. They can have robust designs and enhanced dedicated controllers. However, they have limited processing power and memory, may be larger than the patient would prefer, and still may need a remote device for activation and cloud services.

Location is important. Current subcutaneous systems can advance, but will reach a limitation due to glucose lag and insulin kinetics. Hybrid systems with either the sensor or the pump implanted will do better, but still have the barrier of one of the devices. Implantable systems theoretically may be the best, but have additional challenges. Delivery of insulin into the peritoneum has well-known advantages. There are both IP pumps and IP catheters. Insulin delivery is much faster and lag time for glucose is less. Thus, an IP AGCS may have an advantage over a traditional SC system. IP insulin delivery also seems to be faster, more consistent, less variable, and have smaller swings. The Roche Diaport may be a way to deliver IP insulin without an implanted pump. It connects an external pump to the peritoneum and can be disconnected or replaced as necessary. However, it is more expensive and may lead to very serious peritonitis. In silico, we compared an AGCS delivering insulin SC to one delivering IP, both with unannounced meals. The IP system was far superior with a 50 mg/dl additional reduction in average glucose concentrations.

High performance will require integration of devices and ideas. High performance needs to be built in, not added on. An implanted IP system needs to have methods in place for replacing insulin and sensors when needed. An IP pump needs to be thin, long-lived, and require minimal refills, necessitating the use of concentrated insulin preparation to minimize volume, like Thermalin U1000.

In silico simulations of an implanted system with a PID controller shows outstanding control without announced meals. An implanted controller needs advanced design features: low power consumption, extensive memory, fast duty cycles, great precision, good communication, and it also needs to be resistant to fibrosis by the body and not produce too much heat. MPC can also work as shown by an in silico simulation.

Progress on Novel Insulin Formulations and Delivery. Bruce Frank

Insulin is a very stable molecule, made more stable by the chelation of zinc by the hexamer. In order to increase the speed of insulin delivery we need to get beyond zinc. Humalog is much faster than regular insulin, but is still slow on uptake and has a long tail of activity. Newer AP insulins need to be more rapid in on-set and off-set and, to be useful in smaller devices, also need be both much more concentrated (U500 and U1000) and very stable. Stability is a problem in general (28-42 days), but especially with pumps (3-6 days), since the agitation and temperature tend to cause fibril formation.

To facilitate absorption, insulin needs to be monomeric; but, most monomers are unstable. We can improve insulin action by both accelerating absorption or shortening the duration of action (by abbreviating the tyrosine kinase signaling system). Currently, insulin action goes on for about 3 hours after the insulin is cleared from the blood. This is true for both regular and lispro insulins. We also need the insulin to diffuse into the capillaries faster. With monomers alone, higher concentrations are associated with faster diffusion. However, we have a dilemma: monomers give us speed but hexamers give us stability. In addition, as you concentrate insulin, it forms aggregated hexamers, which are even slower. We need a new approach.

The total time for insulin action is the time for diffusion from the depot, absorption into the capillary, lymphatic absorption, blood circulation, and insulin action at the cell. For lispro this is a total of about 5.5 hours. Proposed solutions include four excipient based solutions: Linjeta (Biodel, suspended company activity), PH20 (Halozyme, suspended company activity), FiAsp (Novo), and Biochaperone lispro (Lilly). Linjeta is monomeric, but none of these shorten the signaling. Thermalin has a new medical entity that is monomeric, has absorption excipients, and shortens signaling.

Linjeta chelates the zinc and charge masks the material for faster capillary absorption but it loses stability. PH20 is based upon using hyaluronidase to break down hyaluronic acid and speed absorption. The safety of long-term injections of hyaluronic acid is unknown. FiAsp uses niacinamide as an excipient to modify absorption and L-arginine to increase stability, but the increase is speed is very little. Biochaperone insulin uses an active polysaccharide to mask the charge and facilitate absorption. It also has a slightly shorter tail of action.

Thermalin is developing new molecular entities, mostly monomeric insulin analogs that are very stable (more than 45 days agitation at 37 degrees C (Celsius). T0339 is a stable analog that is about 8 minutes faster in onset than lispro, even at U1000. We have also identified 14 analogs that have a shortened duration of action (up to 25% faster). Thus, we believe that a stable ultra-rapid insulin can be created that has significantly shorter duration of action.

Progress on Novel Glucagon Formulations and Delivery. Steven Prestrelski

The target product profile for glucagon in an AP is: (1) An indication for prevention / treatment of hypoglycemia; (2) PK of rapid onset and elimination (accomplished); (3) PD of rapid increase in blood glucose levels and rapid loss of effect (accomplished); (4) Total daily dose of < 1 mg; (5) Concentrations of 1-5 mg/ml; (6) No safety issues; (7) Little or no nausea and vomiting; (8) Special vial for transfer to an AP system; and (9) Shelf stability of 2 years and in-use stability of 6-14 days.

Glucagon stability is a challenge. It fibrillates rapidly and is highly susceptible to degradation in water. Currently glucagon is an impure powder that is reconstituted but is not stable in solution. By 8 hours it has lost most of its potency. Thus, current glucagon formulations are not appropriate for a commercial AP.

There are several approaches to improving stability: (1) Add GRAS excipients to stabilize the glucagon (Xeris, Biodel, Latitude, Arecor); (2) Add novel excipients (Curcumin—Ward, Biochaperones—Adocia); and (3) Develop glucagon analogs (Zealand, Lilly, Sanofi). The regulatory risk of (1) is small since it uses native glucagon and well-characterized excipients. It is higher for (2), which uses native glucagon but excipients of uncertain toxicity, and highest for (3), which changes the glucagon molecule and is a new drug.

Xeris (XeriSol) uses a nonaqueous solvent to stabilize glucagon. Biodel uses an aqueous, neutral pH formulation with stabilizing excipients. Arecor (Arestat) uses novel formulation spaces (buffers) to minimize degradation. Latitude (Nano-G) uses an injectable nano-emulsion with glucagon inside micelles.

Xeris loses less than 20% of its potency over 30 months at 25 degrees C. The PK and PD is similar to native glucagon. Biodel gets 90% stability at 10 days at 37 degrees C. The Zealand product is stable for about a month at 40 degrees C, 6 months at 25 degrees C and over a year at 5 degrees C. Thus, a stable glucagon is on the horizon.

Discussion

(Aaron Kowalski) What would patients like to see in AP systems as they develop?

(Alecia Wesner) I have been in 2 AP trials It was amazing to see the stability of the blood glucoses, day by day. Knowing that I and the other people with diabetes can’t use these systems yet outside of clinical trials is a shame. (Adam Brown) I used the DIAS system with control to range for 6 months. I was relieved to give it back. It wasn’t providing good control during the day, but overnight it was great. The connectivity and the alarms were big problems. We need to consider the tradeoffs. (Jennifer Schneider) The overnight control and improved sleep was the main benefit from my experiences, restoring normalcy to the family. (Alecia Wesner) I am an industrial designer and have been in 3 AP trials and no one had asked me about the design or the interface.

(John Pickup) Who should get the AP?

(Jennifer Schneider) These tools are so helpful. I think these technologies do not add complexity. The AP systems should be widely available, even though some people may choose less involved technologies. (Adam Brown) People who prefer injections and BGM generally do so for 2 reasons: They don’t want to wear something on their body and the technology is too expensive. To be widely available, AP systems have to address those two problems or be so effective that people will ignore the problems. (Henry Anhalt) There are still lots of endocrinologists who do not use pumps, will they use an AP? Industrial design is critical and there will be an iterative growth in design and ergonomics. (Robert Vigersky) In the MiniMed 670G trial, 80% of the patients elected to stay on the device. They are providing a lot of posttrial data. Cost and availability are probably going to be the biggest barriers. We need to work with payers with good data to show the benefits.

(Aaron Kowalski) What is the low hanging fruit that will add value to AP?

(Robert Vigersky) The next step is better algorithms followed by personalization of the system. (Vincent Crabtree) We need to think about what is going to make people switch to an AP. We are looking at ultra-concentrated insulins, smaller pumps, and geospatial information.

(John Pickup) Are the medical professionals a barrier to adoption?

(Alecia Wesner) I am my own advocate. I shop for a doctor who will work with me. (Adam Brown) I use an academic site and they still are afraid of the technology. I am annoyed that you cannot “test drive” a pump. You can test drive a car and a pump is way more important to me. (Alecia Wesner) Pumps have poor industrial design. There is no universal interface. If you know one pump, you don’t know how to use another. (Aaron Kowalski) the mobile phone creates a platform that everyone knows. Why not use it? (Stacey Anderson) The phone has replaced my camera, books, mp3 player, calendar, mail, GPS, maps, checkbook, computer, shopping list, and more. Why can’t it effectively replace the controller? (Robert Vigersky) Endocrinologist often do not take care of diabetes or don’t use technology. I have tried to get the Endocrine Society to train our fellows to learn technology. (John Pickup) Should we limit this technology to specialty centers? (Robert Vigersky) Limiting this technology to specialty endocrinologists would be a mistake, but we need to do a better job of training providers. Decision support for both patients and providers is critically important and we are not doing a good job now.

(Aaron Kowalski) We are taking access very seriously. What can we do to improve access?

We need to get remote support for rural individuals, perhaps telemedicine. (Judy Sibayan) Access is a big deal, especially for clinical studies. (Sean Sullivan) At the Trust we have a few initiatives that are addressing the issue of access. For example, we have a rural project to educate primary care physicians; and, in New Mexico, we have a project called EndoEcho to teach primary care physicians to take care of type 1 diabetes. We also have a project with JDRF and the TIDX examining barriers in primary care physicians to prescribing technology. (Dominic Galante) The payer is one of the barriers. You need to provide medical, psychological, and economic arguments for more complex, more expensive devices. (Amanda Bartelme) The biggest problem with payers is not going to be a coverage issue but limits on the population that has access to the technology and how much of the coverage cost is passed onto the patient. Finding creative ways to demonstrate long-term benefits of using AP systems is going to be important. (Dominic Galante) Payers are worried about moving these devices to type 2 diabetes. The cost of this group using higher technology is enormous. (David Klonoff) We need to reach out to endocrinologist and diabetes educators to get them to understand the technology. (Henry Anhalt) TIDX has been working with JDRF to study outcomes of the various devices. Quality improvement will drive this back to the payers. (Joshua Cohen) All endocrinologists at the time of their training need to demonstrate that they be proficient in diabetes. The boards actually test for proficiency in CGMS and insulin pump. We need a way to get that training to those that are already trained. Current payment systems don’t pay for direct contact, but these systems will require a lot of indirect contact (computer, smartphone, telephone) and we need to figure out how to get reimbursed. (Robert Vigersky) I want to address the concerns about costs in using AP systems in patients with type 2 diabetes. Many patients with type 2 diabetes are on multiple expensive drugs and I would suggest that an AP system may provide better control at lower cost. (John Pickup) What is your vision of an AP system for individuals with type 2 diabetes? Can they be simpler? (Robert Vigersky) My guess is that we will use the same algorithms. Individuals with type 2 diabetes tend to be less adherent because they can get away with it in the short run, but do poorly in the long run. An AP system may prevent this. (Steven Russell) There are probably about 1.5 million people with type 2 diabetes who take MDI. They seem to be able to use an insulin only system targeted at 100 without much hypoglycemia. (David Klonoff) The two big problems with people who are doing poorly are generally high HbA1c and/or serious hypoglycemia. But what they like most about systems that bring them back into control is that they feel better.

(Aaron Kowalski) What is the FDA feeling about the state of technology?

(Yiduo Wu) We are excited about the rapid movement to an AP and the first system is under review (now approved). We know that first generation systems will not be perfect. There are challenges to the systems but as the systems come to market we will get better systems with more accurate CGM systems and pumps. (John Pickup) What is the FDA view of the evidence that has been presented to them about AP systems? The clinical trial data are gathered in compliant patients who have previously used Sensor Augmented Pumps, but the likely first candidates for clinical use will be patients with intractable hypoglycemia or elevated HbA1c. (Yiduo Wu) We cannot comment on the data under review. We expect a much larger postmarketing studies in broader patient population for these products. (Stayce Beck) We highly recommend that investigators look at a broad variety of patients in their clinical trials. The risk is that we might limit the group of intended users. (Aaron Kowalski) We are proud of our work with you on the AP guidance document. What is the postmarket pathway you are suggesting? (Stayce Beck) we have been aggressive at getting these systems out to the public but need to have some assurance of safety. We negotiate the premarket and postmarket studies with the manufacturer. We are open to a lot of different ideas.

(David Klonoff) In some ways, AP systems are similar to self-driving cars. Perhaps the automotive industry has some hints for us. (Adam Brown) Automobiles kill 1 million people per year globally and this serves as a baseline risk to which to compare the self-driving cars. We need to understand the baseline risk for diabetes, to which we can compare the AP. (Aaron Kowalski) I want to thank the T1DX for starting to gather this data.