Australasian Science Magazine

The average person with diabetes is outside the optimum range of blood glucose levels more than 60% of the time.

By Jenny Gunton & Nigel Greenwood

Researchers hope that within 3 years new insulin pump software may be available to replace the functions of pancreatic beta-cells lost in Type 1 diabetes.

Type 1 diabetes (T1D) is an autoimmune disease that can be diagnosed at any age but most commonly starts in childhood. In the past it was also called juvenile-onset diabetes, childhood-onset diabetes or insulin-dependent diabetes mellitus. However these names are no longer used because many people with Type 2 diabetes also require insulin, while some children have Type 2 diabetes and some adults develop T1D.

There are thought to be 130,000 people in Australia with T1D, and about 95% of children with diabetes have T1D. The incidence of T1D is increasing by more than 3% every year.

Insulin is a hormone in the blood that keeps blood glucose levels from rising too high by helping glucose enter muscle, liver and other cells. Beta-cells are the only cells in the body that release insulin. They live in islands in the pancreas called islets of Langerhans.

In T1D, the immune system becomes confused and attacks the beta-cells in the pancreas. Without enough insulin, the blood glucose levels (BGL) rise. Before the discovery of insulin in 1921, T1D was a uniformly fatal disease.

Insulin revolutionised the treatment of T1D, but has left the ongoing problem of controlling blood sugars. Because people with T1D cannot make enough insulin they must inject it, either several times a day or via a pump that delivers insulin under the skin continually. In addition, they need to test their blood glucose levels by fingerstick many times every day. BGLs are very difficult for some people to control.

Low BGLs (“hypos”) are associated with unpleasant symptoms like shakes, sweating, nervousness and hunger. Since the brain needs glucose to work, as hypos progress people become confused and then unconscious. Over time, high BGLs and hypos increase the risk of complications that can affect eyesight, the kidneys, feet and increase risk of heart attacks and strokes.

The major goals of diabetes researchers are to improve insulin administration to prevent high BGLs, avoid hypos, improve quality of life and prevent complications of diabetes.

The Artificial Pancreas

In normal people, beta-cells automatically sense how much glucose is in the blood and release the right amount of insulin to control their BGLs. When you eat, especially carbohydrates, blood glucose starts rising and you release more insulin. Normal beta-cells are clever enough to stop releasing insulin as BGLs fall, so that levels don’t become too low.

In T1D, this finely tuned system is broken because most beta-cells are dead. With few beta-cells left you cannot release the amount of insulin needed, so BGLs rise.

Injected insulin does push BGLs down, but it takes a while to act (allowing high BGLs). Once injected, insulin keeps acting for 4–48 hours regardless of whether BGLs still need to go down.

One way researchers are trying to help people with T1D is by attempting to develop an artificial pancreas. The major goal is to combine glucose test results with insulin infusion so that glucose tests “talk” to the insulin pump and tell it what to do.

This is a lot harder than it sounds because the processes that cause glucose levels to change are really complicated (Fig. 1). Some things can make BGLs rise or fall depending on time and the individual.

Figure 1. Factors causing high or low blood glucose levels (BGL). High BGLs are levels above 6 mmol/L before breakfast or above 8 mmol/L 2 hours after meals. Diabetics often have BGLs in the high teens. This increases the risk of complications, and can give symptoms like thirst, excessive urination, and feeling tired or mentally “dull”. Low BGLs are common for most people with T1D and can be dangerous. The brain needs glucose to work so if BGLs become too low, the brain stops working properly. This eventually leads to confusion and then loss of consciousness.

A way to solve this is to try to design software smart enough to reconstruct what the normal body does. It would need to interpret what the BGL results mean at a particular moment (e.g. whether BGLs are going up or down, and how fast) and then calculate an appropriate insulin dose to give and what pattern to give it in.

There is currently a worldwide race among groups of scientists to design artificial pancreas software that will work reliably and safely.

A New Approach

The key problem is the amount of uncertainty surrounding BGLs due to the variety of factors that can influence them (Fig. 1). The second problem is getting enough information. BGLs can be measured using fingerstick tests, but four measurements per day doesn’t give enough information to program most software.

A surrogate marker for BGLs is to measure glucose between cells via a tiny plastic tube inserted under the skin. Interstitial fluid glucose (ISFG) can be measured every few seconds using a continuous glucose monitor (CGM).

However, this brings in a third problem. Glucose takes time to get out of the blood into the interstitial space. How long varies from person to person and from time to time. This can be really important: if you wait until CGM detects a hypo, the BGL may have been low for 10–20 minutes!

Because of the many variables involved, conventional forms of automation don’t work well for an artificial pancreas. New approaches are needed to make it more reliable, safe and stable.

In our collaboration, a new approach has been taken using “machine-intelligent” software, a form of artificial intelligence. Good machine-intelligent software is slightly paranoid and anxious. To borrow a line from Donald Rumsfeld, it has “known knowns” (i.e. it “knows” verified rules about the world around it), “known unknowns” (i.e. it knows that some things cannot be known accurately but will be within known boundaries) and it is aware of “unknown unknowns” (e.g. its world-view is constructed from imperfect mathematical models and it doesn’t know what these imperfections are; or the models may be accurate but the input data has some false entries).

The last two explain why well-designed machine-intelligent software is “slightly paranoid and anxious” – the software must constantly reassess its world-view as new data comes in, look for inconsistencies and contemplate worst-case scenarios.

In our version of the artificial pancreas, software was written to deal with all sorts of potential errors and changes, including the lag between ISFG and BGLs, the ongoing effects of previous insulin, and the risk that current data suggests an impending hypo. It was also designed to handle a major source of real-world uncertainty for people with diabetics – meals.

In real life, most people don’t eat meals at precisely the same time each day and eat the same meals, so their carbohydrate intake varies. Most people cannot accurately calculate how many carbohydrates their meal contains. At work, many people do not know when they can take a meal break more than 10 minutes ahead of time. So, the software had to be written to not be sure how many carbs were being eaten, or that the meal was coming more than 10 minutes before it started.

Being able to handle all this uncertainty takes a lot of computing power. Fortunately we have two important resources. Smartphones are actually sophisticated microcomputers, and cheap supercomputing resources have become available via the internet for intermittent use.

Of course, any software for T1D must still work if internet connections go down.

Initial Tests

We carried out the initial tests of the software on simulated patients using data from the medical histories of people on the pancreatic islet transplant waiting list. To get on this waiting list people must have T1D, severe hypos, frequent highs and difficulty controlling them. By combining medical histories with already published models of T1D, we created simulations of two diabetic patients. Using actual insulin infusions, BGL, ISFG and meal data from people’s records meant that assumptions could be avoided and we could be confident that the simulation tests were as realistic as possible.

Our artificial pancreas software used a high-performance computer to reconstruct personalised models of the BGL processes using a day of real data from these simulated patients. For each patient, the software then verified the validity of these personalised models by taking actual insulin and meal data from 8 more days, predicting BGLs and ISFG levels, and comparing the predictions with the actual results.

We found that the models were extremely accurate when fasting (before breakfast), with some uncertainty emerging once meals began for the day.

Once it was satisfied with this “validity-testing” it was necessary to help the machine-intelligent software understand what it was observing. To do this it generated 2599 possible diabetic profiles to study and control. Out of the maze of outcomes the software suggested insulin dosages that were given to the simulated patients in a “closed loop”– insulin was applied throughout the day using an entirely automated process to control the simulated pump, while the ISFG and BGL measurements continued to be taken as if the simulated patient was going through a typical day, with the ongoing readings being put back into the software to reassess insulin dosage.

Over 55 generated 24-hour days, the results were extra­ordinary: no hypos or any severe post-meal highs. BGLs were kept within the target interval for more than 90% of the time, and the target was achieved 99.8% of the time in the main case studies. This is extraordinary because the average person with diabetes is outside that range more than 60% of the time (i.e. they achieve the desired target less than 40% of the time).

Figure 2 shows an insulin profile commanded by the software after reconstructing a personalised model, and Figure 3 shows the consequent effect on the simulated patient.

What Does It Mean?

These early results are exciting, but much more is needed. The next step is to repeat this experiment in real people rather than through a simulation. This would initially involve a small clinical trial using volunteers with T1D. If it works as well as we hope, a large randomised clinical trial will be needed with many more patients.

We hope to achieve regulatory approval in 2016 so that the software can be made available to people with diabetes by late 2016.