Digital ecosystems for managing diabetes

The needs of people living with Type I diabetes

Diabetes is often talked about as a “growing epidemic”, however it’s important to distinguish between Type I and Type II diabetes. Type II diabetes is simply an increase in insulin resistance, caused by age, being overweight or both, which can often be treated by common-sense changes to diet and lifestyle. When we hear about the diabetes epidemic, it is normally this type that is being referred to.

Type I diabetes, on the other hand, means the body produces little or no insulin on its own, relying on extraneous delivery of this vital hormone. This mode of delivery causes a whole host of complexity for Type I diabetics but cannot yet be avoided. Whether delivered by insulin pumps or insulin pens, taking doses as recommended and tracking progress over time can create a host of complexities for users.

Fragmented digital ecosystem for managing diabetes

As mentioned, there are currently a number of condition management systems for Type I diabetics however, these are often on separate platforms and devices which are unable to communicate with each other. For example, while the insulin pump and continuous glucose monitoring (CGM) system I use are integrated and allow me to check the current active insulin and serum glucose status on my phone, viewing reports over time must happen on a laptop. This pump therapy progress and reports are also not necessarily available to my GP, which leaves them without a key data point for any decision-making. Meanwhile, calculating carbs, recording meals and recording exercise must be performed using other services which are not linked to my pump. The same goes for any training I receive and associated digital services, such as DAFNE.

Managing all of this, on top of managing my daily condition, can be a headache, figuratively as well as metaphorically.

Challenges with managing exercise

Managing exercise provides one of the most complex challenges for diabetics. For example, aerobic exercise such as running may seem like a simple activity. However, a 30-minute run vs a 45-minute run may require entirely different pre-run dosage adjustments, varied carbohydrate intake while running and changes to recovery schedules and management of background insulin in the evening. Anaerobic exercise, such as lifting weights at the gym, is also fraught with complexity. Anaerobic exercise is technically supposed to raise blood glucose in the short term, however many diabetics who attend the gym regularly will tell you that it isn’t this simple. Depending on the length of workout and types of exercise, you may be performing a mix of aerobic and anaerobic exercise. How should you adjust your dosage routine before, during and after?

There are also challenges with using insulin pumps when cycling, long distance running or other activities which may last for a number of hours. Even with preparation ahead of time and setting higher serum glucose targets on my pump, I will regularly find that hypoglycaemia is fast approaching when exercising for more than an hour, even when taking in carbs. The pump will also shut down its intelligent insulin delivery system if insulin delivery has not been required for a few hours. This can be massively frustrating when trying to recover after a long ride and having no automatic insulin adjustment for up to five hours.

Ideally, my system would give me the ability to input an exercise type, expected duration and possibly intensity (although that can be a little subjective). This ideal system would automatically calculate and make adjustments to insulin delivery before, during and after exercise, as well as suggesting when to take in carbohydrate to maintain energy and avoid hypoglycaemia.

While there are already a wide range of wearable devices for tracking fitness and exercise, including but not limited to FitBit, Oura, Apple Watch and StatSports vests, none of these products currently communicate with my insulin pump and are not medically certified devices. While this creates a barrier for including any of these devices in a treatment or diagnosis workflow, thankfully Garmin are clinically trialling at least one of their wearable devices for medical certification of ECG capabilities, so watch this space.

Carb-counting and eating

A well-known challenge for diabetics is carb-counting, which involves calculating the amount of carbohydrate in a meal, often based on a visual estimation. Services such as Carbs & Cals are invaluable for this purpose, providing visual representations of various serving sizes for a particular type of food, with the suggested carbohydrate amount listed. SnapCalorie and SNAQ go one step further in easing dietary monitoring for the user, by allowing an image to be taken of a meal and then offering the user suggested nutritional values based on it.

An added complexity for diabetics can come from managing protein intake. For example, if you eat a chicken salad which doesn’t contain any carbohydrate directly, the body will likely then convert some of the protein into carbohydrate over the next few hours. This adds unpredictability to the carb-counting and dosage timing.

There is also the challenge of knowing when to take an insulin dose. When using a pump or a pen, it is generally advised to take insulin 20 minutes before eating. This means you need to estimate the exact amount of carbohydrate you’re going to be eating 20 minutes before you begin eating. This is often problematic, for example, if you don’t eat as much as intended, are eating out and waiting for food to be served, are visiting friends or relatives or if you simply forget to check the pack before cooking yourself. If you take the insulin too late, you see a spike where your serum or blood glucose goes above range and then a pump may over-compensate for this, causing hypoglycemia three hours later. It really is a minefield.

Viewing status and data visualisation

Viewing live status, or near to live status, of serum glucose is becoming very much the norm for Type 1 diabetics. Devices like the Freestyle Libre have paved the way for much more robust and meaningful treatment tracking. Gone are the days of a GP asking “So how much insulin do you take a day?” or your care team examining your glucose meter readings for a week. Now, the vital measure of “time in target range” is the gold standard for progress measurement in treating diabetes, with the HbA1c blood test being the clinical backup for this measure.

Data visualisation is, however, a mixed bag. Abbott include up to 90 days of data in the Libre app, which can be hugely useful for tracking over time. The Dexcom app offers a similar ability, although the data visualisation is a little clearer and more extensible in the Libre app. The insulin pump I am using has a proprietary data visualisation platform which is unfortunately not very user friendly. It seems like there is still room for improvement in the user experience of these devices. In the past, data visualisation for diabetes was often aimed at clinical teams. In a number of cases, these clinical graphs and reports are now accessible to patients via their apps. As we discovered in a project we conducted at Team, lab reports can be quite confusing and we can do a lot better to aid patients in visualising their data.

Connected insulin delivery devices

Whether using an insulin pump or pen, CGM systems, flash glucose monitoring or finger pricks, many devices are built with connectivity in mind. In a recent development, the NovoPen Echo system can be synced with the Libre system. Now patients can tap the back of their phone with the pen to sync their insulin doses with the Libre’s glucose measurements. This helps to give a much more accurate picture over time, with less room for human error and less overhead on the user.

Pumps and CGM systems can often be used as a hybrid closed loop system, sometimes referred to as the “artificial pancreas” approach. Systems like CamAPS, Medtronic’s SmartGuard and the Ypsopump myLife system offer this functionality in a variety of ways. CamAPS and myLife use the Dexcom G6 CGM while Medtronic use their own proprietary Guardian sensor.

While these systems are great at what they do and can offer clinically significant improvements for patients, regarding time in optimal blood glucose range, their inputs are limited to CGM sensors alone. None of these systems currently utilise input from heart rate monitors or other medically certified connected biometric measurement devices, which can mean managing exercise and illness are no less complicated than when using injection pens.

My ‘dream system’ for managing diabetes

With continued advances in machine learning and wearable devices, we are not far from a reality where a smart assistant could be used to help track and manage diabetes in “real-time”. I like to think of my perfect workday where I enjoy my life to the fullest, free of hindrance, using such technology. It would look something like this:

• Waking up: I wake to my smart assistant, Ali, who asks me what I think I’ll be eating for breakfast, based on my regular choices. The system starts to deliver the correct amount of insulin to prepare for this as I make my way to the kitchen.
• Commuting: I’ve told Ali I’ll be cycling to work, so my insulin dosage has been automatically adjusted based on this exercise. I use a number of connected devices such as a Fitbit and Garmin GPS tracker, the data from which Ali uses to make better predictions about my dose.
• At work: At around 11am Ali tells me I should eat a snack based on my blood glucose level. At lunchtime, my colleagues ask if I’d like to join them at a local Japanese restaurant. I say yes, absolutely. When ordering, I think I’ll have three different types of sushi and inform Ali. I’m given a conservative insulin dose a few minutes later, before receiving a full dose when I go on to tell Ali I’ll be finishing the lot.
• After work: A friend suggests a game of squash, which means I’ll have to have dinner a little later to accommodate, but that’s OK. Ali will be prepared to reduce my background insulin in preparation for playing. When I get to the court, I need to remove my pump and leave my phone on the table next to it, so I’m reminded to take a glucose gel before playing. After 20 minutes, Ali uses data from my connected heart rate monitor to calculate that I’m going to go low before my activity finishes and advises me to have a second glucose gel.
• Dinner: When I get home, I log my protein shake with Ali and decide I’m going to have a chicken salad for dinner as I’ve had enough carbohydrates for the day. Ali knows that some of the chicken proteins may be broken down into glucose a couple of hours later and calculates when to give me the correct dose of insulin based on a photo I take of my meal.
• End of day: Ali flags that my serum glucose is dropping after dinner and I may have a hypo within half an hour due to the earlier exercise, so I’m advised to have a few biscuits or some fruit. Before I go to bed, I notify Ali that I’m going to be helping a friend build his shed tomorrow afternoon for a few hours. My insulin dosage is adjusted accordingly for the next 18 hours and I go to sleep happy that I have a helping hand in life to manage my condition.

How far away is this future?

Normally, a projection like the above would be accompanied with a statement such as “in 5-10 years, all of this could be possible”, but in reality a lot of this could be nearer than we think. Systems such as ChatGPT and Replika, both powered by AI and available right now for consumer use, can already be used to make natural language interactions like the above a reality.

The challenge will come with the clinical provability of these interactions and may take a leap of faith from manufacturers and regulators to enable systems like this to make suggestions related to treatment. AstraZeneca are already working with regulators on how to garner approval for machine learning and AI systems, involving regulators in more detail at every step of the process of software development. In the short term it seems this approach will be needed by everyone with an interest in this space. However, there are already resource constraints within regulatory organisations, given the huge numbers of digital products they are being asked to approve. A different approach may therefore be needed. The FDA have a plan to develop a framework around AI-based products and services for healthcare, so hopefully other regulatory bodies will follow suit.

Medical device manufacturers provide life-changing products but often resist suggesting changes to the treatment of diabetes with their systems. Human interpretation by clinical care teams is almost universally required when considering a change to treatment, or to prompt a change in patient behaviour should persistent problems occur. The most scalable solution and more likely the most beneficial solution for patients, will be an AI system that can react instantly to problematic behaviours or emergency situations.

Managing exercise seems to be one of the most complex and often missing variables in diabetic management systems. Machine learning systems, which can be simpler and easier to build than AI systems, could be employed to manage this aspect. There are a great number of tracking and exercise management services for athletes that use intelligent approaches, so there may be a way to leverage these. It’s not hard to imagine the impact that could come from combining the intelligence of apps such as Noom, SnapCalorie and Strava, with wearables like those provided by Catapult, Garmin and more. Such a combination could provide an integrated diet and exercise service, although these elements would need to be linked symbiotically to stand the greatest chances of success. Add an AI interface that both sends and receives natural language requests and you have a fully functioning, digital ecosystem for managing diabetes.

There are obvious challenges in mixing non-medically certified apps and services from the wellness space with medical treatments and diagnostics. While there is already precedent for separating Software as a Medical Device (SaMD) from non-medical software when developing health technology, this generally covers development of software from scratch. It may be that the first instances of a super-service combining multiple inputs and services are not medically regulated, which could be problematic. The hope is that regulatory bodies are aware of this and are working to adapt their processes and policies to be streamlined for the ever-growing number of digital health products being submitted for approval.