A robot and a human clink wine glassesFor most of us, AI is the sound we make when we find out a robot has taken our job.

For Professor Chris Gilfillan and his team of researchers, Artificial Intelligence suggests a convenient way to provide personalised advice to diabetics.


a robot chops up some tomatoesKey Ingredients

We’re all besieged by well-meaning dietary advice every day, from the best type of lard to deep-fry your butter in, to the worst time of day to avoid celery. As people with type 2 diabetes are often encouraged to modify their lifestyle to help manage their condition, a cornucopia of apps, books, podcasts and websites have burst forth promising to help them manage their blood glucose levels.

However, glucose response is highly idiosyncratic, and the type of food that causes a blood sugar spike in one person might not register a blip on another’s glucose monitor. Existing diabetic guidance represents the most comprehensive dietary information we have at a population level, but it’s generalised data that often doesn’t correspond to a person’s individual experience. No diabetic body aligns perfectly with the generic body of data. Any good diet book will recommend replacing white bread with brown, but an individual might find that brown bread raises their blood sugar level just as high and makes lacklustre fairy bread.

This can not only be confusing, but frustrating and demoralising. A diabetic person who feels like they’re doing all the right things but their body isn’t responding the way experts say it should might give up trying to follow the guidelines altogether. As one participant in Chris’ study stated, “I’ve been diabetic for a bloody long time. Twenty-odd years. And I still don’t get it … I just find the whole process damn confusing.”

a robot scoops flour from a containerMethod

Working with researchers at the Monash Institute of Medical Engineering (MIME) and the Eastern Clinical Research Unit (ECRU), Chris Gilfillan developed an app that would provide personal feedback for individuals after learning about their specific reaction to particular foods.

The app, named GLOOK! asked participants to take photos of their meals while collecting biomedical data from their glucose monitors and Apple Watches. The data would then be compared against the dish to link ailments with aliments.

GLOOK! was originally based on raw biological intelligence, with photographs and data being received and reviewed by Chris himself. As Director of Endocrinology at Eastern Health, Clinical Professor at Monash and Deakin Universities and formerly head of endocrinology at Frankston Hospital CHECK THIS, Chris has seen more HbA1c than Willy Wonka. At the end of each day, he would compare each person’s meal to their blood glucose levels, analyse the correlations, and send out personalised summaries.

The second stage of this project set out to see whether an artificial intelligence could be trained to replace Chris. The program was first taught to recognise images of food, then fed the participants’ photos and data. This phase set out to investigate whether an AI using the same information as Chris would come to the same conclusions. As the AI was brought in after the original experiment had finished, and the aim was to compare it to a human doctor rather than influence clinical outcomes, it did not send messages to any participants.

a robot pours a cup of flour into a bowlThe Taste Test

In the human-led phase of the project, participants received friendly messages like “Excellent morning after low carb breakfast. The coatings of schnitzels are a trap as they contain rapidly absorbed carbohydrate. Steamed chicken breast may have been a better choice. Well done for the extra exercise on the bike. Although exercise can acutely put the blood sugar up, the overall effect will be positive.” These gentle texts synthesised warm commendations, bite-sized advice, and informative tidbits.

The experiment resulted in a significant reduction of 0.22% in the participants’ average hemoglobin A1c level, a marker for excessive blood sugar. If the best and worst HbA1c results are omitted (representing participants who had extremely well-controlled and poorly-controlled diabetes) the reduction is even more dramatic. Fructosamine and lipid fractions were also affected, but not to a statistically significant extent. Furthermore, 87% of the participants gave positive feedback about the project.

a robot cracks an egg into a bowlThe Bitter Aftertaste

While the first phase of the project was a piece of cake, the second phase bit off more than it could chew.

As the experiment only ran for 12 days, the AI had limited data to draw upon. When given all the photos of food and all the participants’ biometric data, it couldn’t accurately predict which foods affected which people in which way. It was trained on group data, which meant it couldn’t give the personalised, individual recommendations the researchers were after.

When the AI was retrained just on the data of single individuals it did better, but still wasn’t a very good Chris. It also had some trouble identifying the meals the participants photographed, performing better when faced with an uncomplicated dish like a bowl of cornflakes than a convoluted meal like a sandwich.

A factor in the AI’s inefficacy was the messiness of real-world human behaviour. Blood sugar levels are affected by all sorts of things other than food, including exercise, stress, and sleep. After rising, they take hours to settle down.

Ideally, participants in a study linking food to blood sugar results would eat one meal a day consisting of a single ingredient in order to establish a clear cause and effect. Instead, participants in this study behaved like normal humans, eating meals made up of many ingredients, snacking multiple times a day, exercising inconsistently and feeling emotions. This made it difficult for the AI to associate specific foods with a specific blood sugar result.

A robot squeezes a lemon into a bowlTurning Lemons into Lemonade

The failure of the AI taught valuable lessons about the importance of thorough training and the necessity of good data quality and quantity. Without a good teacher, the AI could not learn enough to give meaningful results.

As AI seems like it will become an inevitable tool in healthcare settings, experiments like this help to set future developers up for success. Looking at the ways and the reasons the program under-performed will be useful for other AI developers to avoid the same pitfalls.

Eventually, Chris and his researchers hope to release an AI-enhanced GLOOK! to help diabetics, pre-diabetics and health-conscious people learn about their own bodies. Daily personal dietary advice from an award-winning endocrinologist isn’t something currently offered to diabetic people, but GLOOK! could deliver the next best thing.

a tasty cakeA recipe for success

Chris’ advice for other researchers is simple – if you have an idea about how something could be done better, do it! Don’t be afraid to propose wild research ideas, as unique perspectives and novel approaches have the potential to uncover new insights and develop better methods. Write up a proposal, start looking for funding, and don’t give up at the first hurdle. While research has its unpredictable ups and downs, like blood glucose levels, a positive trend can be established over time with persistence and patience.