Why machine learning is the future of treating diabetes, one of the fastest-growing health crises of our times
“Diabetes is one of the fastest-growing health challenges of the 21st century,” according to the International Diabetes Federation (IDF), “with the number of adults living with diabetes having more than tripled over the past 20 years.” 90% of those with diabetes have Type-2 diabetes, while 8% have Type-1. Now MachineGenes is here to help.
By effectively building an AI pancreas – or at least, a model for the patient’s pancreas, based on their medical history – MachineGenes is able to create a tailored insulin strategy for the patient for the day ahead. For people with highly unstable glucose levels or complex insulin-glucose dynamics, this can allow for precision dosing, and can potentially allow them to go on with their lives as normal, and reduce their risk of diabetic ketoacidosis.
Nigel Greenwood PhD, Leader of Team MachineGenes goes into more detail, and explains the possible links between COVID-19 and diabetes:
Tell us more about MachineGenes.
We’re based in Brisbane and Sydney, Australia. Algorithm design, programming, and general administration are in Brisbane; clinical work is done in Sydney (Westmead Hospital). The team has six members. Brisbane: I’m the team leader (mathematician and architect of the AI algorithms), then there’s Dr. Ingo Jahn (University of Queensland engineering lecturer, who originally came on board to help our team with another project and stayed for our diabetes work), Dr. Brruntha Sundaram and Mr. Alex Muirhead (programmers). In Sydney, we have Professor Jenny Gunton and Associate Professor Jane Holmes-Walker (two of Australia’s senior endocrinologists, at Westmead Hospital and the University of Sydney).
What's the global issue you're tackling as part of the $5M IBM Watson AI XPRIZE?
In brief, we’re transforming the quality of insulin dosing for people with insulin-dependent diabetes, most obviously Type-1 diabetes (T1D), so that people with even the most unstable forms of T1D will be able to have safe, stable, effective doses of insulin, transforming their quality of life. Although T1D is our main focus, other people who will benefit from our work are people with Type-2 diabetes (T2D) who may also need to use insulin, and people with Type 3c diabetes (T3cD).
Insulin is a hormone that regulates blood glucose (BG) levels in the human body. T1D is an auto-immune disorder that kills all the insulin-producing cells in the body. They never grow back, which means that people with T1D must take medical insulin daily for the rest of their lives, to regulate their BG levels. Without medical insulin, T1D is always fatal. T1D is a global pandemic, increasing steeply.
How prevalent is it?
Among adults (over 20 years old) there are currently about 424.9 million people with diabetes worldwide (one in eleven adults), costing $850bn annually in healthcare. This is anticipated to rise to 629 million people with diabetes by 2045, costing $958bn annually in healthcare. In North America 45.9 million adults have diabetes, with over 86 million more pre-diabetic (hence the endocrinologists’ nickname of “the tsunami” for the current global diabetes crisis).
People with T1D, or insulin-using T2D/T3cD need to inject or infuse medical insulin to help control their BG levels. But the problem is, the dynamics of the diabetic interactions between BG and medical insulin are very complicated and nonlinear. Although the general structure of how medical insulin and blood glucose interact is well understood, to understand and predict these interactions to dose insulin properly for an individual person involves understanding their disease “parameters” – a large string of numbers that is individual to each person with T1D, almost as distinctive as a fingerprint.
T1D presents in a very different way across the population of people with diabetes, due to these differences in the parameters. And you need to know these parameters to dose insulin appropriately for someone. At the moment it’s impossible to know a person’s diabetic parameters, so you have to guess the appropriate insulin dose, or work empirically from past experience.
Once you infuse or inject insulin into somebody, it can continue to act for hours within their body, so you need to get it right. To make matters worse, some of the parameters are known to shift in their value across the 24-hour day (Circadian rhythm) or across the month (menstrual cycle), or change according to illness or exercise. So you need something more sophisticated than a human endocrinologist to design an insulin strategy robust again these variations, yet safe, stable, and effective. Enter AI!
So how does your AI technology work?
We have designed, first, a form of evolutionary machine learning, whereby the best models of T1D in the medical literature are combined into simulated “chromosomes.” Confronted with small segments of medical data taken from an individual’s medical history, these chromosomes then evolve, in a simulated environment using high-performance computing, until the models match the medical data. They are then tested against a new segment of the medical history to confirm they remain consistent with that individual’s form of T1D. This is organ-grade personalized medicine, achieved using AI instead of genomics.
But quite a bit of ambiguity remains. So, second, we invented a form of adversarial AI (demonstrated by me in simulation for T1D two years before the invention of Generative Adversarial Networks – GANs – and patented by me six months before the unveiling of GANs by their Canadian team in December 2014) that can interrogate the ambiguities in these evolved chromosomes to build new strategies and simulation-test their safety and effectiveness, before recommending the dosing strategy. All of the steps in this process are fully explainable to a clinician, and the strategies can then be interactively modified by the AI in near-real-time.
These are world firsts. A US patent was granted for this work in February 2020, when I was in New York for the $5M IBM Watson AI XPRIZE semifinals.
Why is what your team is doing important now, and how do you see it scaling up in the future?
Before the onset of COVID-19, the “tsunami” of the diabetes pandemic was a looming global health catastrophe, especially the inability to dose insulin appropriately for those most in need. Once effective vaccines have been deployed to address the current tragic catastrophe of COVID-19, the tsunami of the diabetes pandemic will still be there, and be even more imminent. And so what we are doing remains important.
Is there a relationship between diabetes and COVID-19?
There appear to be serious linkages between COVID-19 and T1D. An editorial published by The Lancet in October 2020 has highlighted the increased likelihood of people with T1D to die from COVID-19, compared with healthy people. It comments that:
“Importantly, the study also shows that the risk of COVID-19-related mortality is significantly and independently related to hyperglycemia in people with either type of diabetes. Hyperglycaemia can impair host defenses, and poor glycaemic control has been associated with infections. Given that glycaemic control is a modifiable factor and can be achieved and sustained by health-care interventions, these results emphasize the importance of supporting people with diabetes in effective self-management. As the global COVID-19 pandemic continues to evolve, it has also become clear that the interplay between COVID-19 and diabetes entails complex pathophysiology. Not only are COVID-19 outcomes more severe in people with diabetes and metabolic dysfunction, but recent data also suggest that COVID-19 could precipitate acute metabolic complications of diabetes, such as diabetic ketoacidosis and hyperglycemia.”
So COVID-19 makes our work on AI-enhanced glycaemic control (i.e. control of diabetic blood glucose) even more urgent. The Lancet editorial goes onto explain that survivors from COVID-19 may be more vulnerable to T1D as a subsequent disease, as well as COVID-19 making our work more urgent in helping people with diabetes to survive.
What team accomplishment are you most proud of over the course of being part of the $5M IBM Watson AI XPRIZE?
Much of the preparatory work building our technology was done prior to the AI XPRIZE. The AI XPRIZE work then involved data-mining actual medical histories of [people with highly unstable T1D from Westmead Hospital and demonstrating our capabilities as applied to real people.
However, the team accomplishment of which I’m most proud is when we demonstrated aspects of our medical AI technology and gradually realized how radical our tech was when compared with conventional AI solutions on the world stage. That was a revelation! (It was the AI XPRIZE that encouraged participation in the UN’s “AI for Good” conference in Geneva 2019, and at NeurIPS in Canada in 2018-2019, and provided a sufficiently high-profile event for us to persuade sponsors to stump up airfares.) It cheered up my programmers enormously!
Outside of your work, what's an area of AI that's exciting you right now?
Outside our work for the AI XPRIZE, but within our other research, is that we’re working on forms of “self-aware” AI for operating systems of complex vehicles, and also on evolving and growing computational models of aviation engines as if they were living organisms.
Completely outside our own research, we’re interested in work being done on what it means for ANN to “dream” and wondering if there’s a way to make our own forms of ML dream as well!