Artificial intelligence in diabetes: The pancreas finally hacked?

Dr. Joris Galland, an internist and passionate about new technologies, discusses the advantages of using artificial intelligence to treat diabetes.

Joris Galland is a specialist in internal medicine. After working at the Lariboisière Hospital (AP-HP), he joined the Bourg-en-Bresse Hospital. Passionate about new technologies, he proposes in our blog "Connexion(s)" to explain the issues at stake.

9am. Like every morning, I attend the transmissions. I'm lucky today, the nurses explain to me that all the patients are doing "quite well". However, I realise - once again - that the morning blood sugar levels of diabetic patients are not balanced! One can be an internist, a cardiologist, a surgeon, a general practitioner... Glycemic control in diabetic patients is a recurrent, complex and annoying problem.

Mr A., for whom I increased the dose of slow insulin last night, was hypoglycemic this morning. Mrs V., who has been taking doses “fit for a horse”, is still hyperglycemic this morning. Her neighbour too: She has been in hyperglycaemia since I lowered the rapid insulin because of her hypoglycaemia. It's enough to make you tear your hair out! The worst thing is that these blood sugar issues prolong the hospital stay, which is not pleasant for the patient or for the hospital's finances.

When I was a doctor at Lariboisière Hospital, it was simpler. There was Jean-Philippe. The diabetologists were all on the same floor, so all I had to do was have a little coffee with Jean-Philippe to get his precious advice on adapting treatments. It's more difficult here in Bourg-en-Bresse: I can't see myself calling the endocrinologists at the other end of the hospital every five minutes - and the hospital is very big. So I did what I know best: In the absence of Jean-Philippe, I looked for a solution in new technologies.

Dynamic measurements for better blood sugar control

Glycemic control, as taught in medical school, is based on measuring capillary blood glucose three times a day. Based on these three results - and taking into account other parameters such as diet - a dose of insulin is prescribed. In view of current technology, I find this procedure completely archaic: These measurements are "static", invasive, painful and probably not very representative of real blood sugar levels over 24 hours. And yet, everything depends on them!

Fortunately, new technologies now make it possible to take continuous measurements, in a non-invasive or minimally invasive, and in any case painless way. In a previous article, I already told you about connected watches, capable of taking continuous measurements. Can you see what I mean? Blood sugar measurements = data = AI (artificial intelligence).

If an AI can beat a human brain at Go, it could very well beat a pancreas - an organ that is particularly poor in neurons and thoughts or algorithms. This is what is about to be achieved with the artificial pancreas. When I first heard about it, I imagined a small electronic organ implanted in the patient's abdomen that would replace the pancreas by taking dynamic measurements of blood sugar levels and then delivering the right doses of insulin through an algorithm. After studying the subject, I realised that it was not a question of 'grafting' a 'robot pancreas'. However, I was right about how this artificial pancreas would work.  

Closing the loop

The first artificial pancreas arrived in the 2000s. Patients wore a sensor (the blood sugar monitor) that measured interstitial blood sugar levels more than three times a day without needles. On the basis of the data received, the patient activated the insulin pump (which he also wore) and decided on the doses to be injected. This artificial pancreas was called an "open" circuit, since the insulin pump was open to the patient's decisions. The device, although innovative, was not so revolutionary because the decision about insulin doses remained dependent on the patient.

This control loop is gradually "closing". Insulin pumps are being minimised, blood glucose watches can act as monitors. The circuit was initially "semi-closed", with insulin injection becoming automatic according to interstitial glucose measurements taken every five minutes. This system interrupts the pump if there is a risk of hypoglycaemia, and the patient can "take control" at any time, i.e. suspend the software. Disadvantage: A "semi-closed" circuit does not exempt the patient from performing capillary blood glucose tests, either to calibrate the device (twice a day) or to confirm a detected hypoglycaemia.  

Since 2011, advances in artificial intelligence have made it possible to close the control loop a little further. AIs can monitor blood sugar balance via an application on a smartphone connected to the pump/sensor via Bluetooth. The calculation of the doses to be delivered is constantly updated and more refined, because the algorithm takes into account past blood sugar cycles and the patient's diet. At the beginning, the patient informs the system of the carbohydrate intake of his standard meals.1 Another advantage is that the patient no longer has to constantly react to alarms and make decisions.  

The DBLG1 device2 was tested under real-life conditions against an "open" system, and analyses over 12 weeks showed that the closed AI-based system kept the patient's blood glucose levels within the therapeutic target range (70-180 mg/dL) for a longer period of time (10-15% longer, i.e. more than two hours per day).3 Another advantage is that the continuous measurement of interstitial glucose provides a weekly trend. This allows the patient and the physician to fine-tune the blood glucose profile.    

AIs alone at the top?

Although the technology is there, the loop is not yet completely "closed". Some diabetologists prefer to talk about "hybrid" systems because the patient still has to inform the system when a meal or physical activity begins. Moreover, one unknown factor remains: What about the role of emotions? We know that a stress spike can have - not systematically - an immediate hyperglycemic effect. Will AI be able to react quickly to an event that is by nature unpredictable? Or will it require human intervention?   

AIs are making great strides, but it is human factors that are still holding back the development of new technologies in the field of diabetes. It is not so easy to entrust one's blood sugar balance to an algorithm, especially if one is not familiar with new technologies! Reassurance about the security of the system, explaining how best to use it... these measures of therapeutic education will therefore be the key - as is often the case with diabetic patients.  

Another obstacle is the constraints linked to insulin pumps, which are cumbersome or have limited autonomy (three days for a "patch" type pump). But if the equipment only needs to be miniaturised and optimised, the algorithms are there! The pancreas, hacked by AI?

Far from insulin pumps, applications are already being used routinely to help patients manage their insulin. Again based on deep learning, these applications use AI to help the patient choose the best dose. The Diabilive application calculates the right insulin dose based on past measurements. It stores the data in order to compile it for the diabetologist. Technology at the service of patient monitoring!

Retinopathies in the eye of AI

Finally, I couldn't end this article without mentioning the IDx-DR algorithm. This AI is able to read the fundus of patients' eyes, without human help. Diabetic retinopathy, a frequent and serious complication of poorly controlled diabetes, can lead to complete blindness. However, everyone knows that it is sometimes long and tedious to get an appointment with an ophthalmologist (even if this situation seems to be improving).

Screening for diabetic retinopathy is an emergency and the USA has understood this. The Food and Drug Administration approved the use of IDx-DR in 2018. This is a first for a diagnostic artificial intelligence system. In practice, patients no longer need to consult an ophthalmologist: They go to practices or booths equipped with an ophthalmoscope and this algorithm.

The AI is able to recognise the signs of retinopathy on a fundus photo with an accuracy of over 90%. If the AI judges the result to be normal, the patient can go home and return the following year. If not, they are sent directly to an ophthalmologist. This saves time for the patient and for the health professional, not to mention the financial implications.

Jean-Philippe vs. AI

I see two solutions for managing diabetes. Have your coffee with Jean-Philippe, or rely on AI to hack the pancreas. I invite diabetic patient associations to support start-ups that are taking on this formidable challenge: Replacing a defective pancreas with increasingly precise and relevant algorithms, to make a diabetic's quality of life as good as it should be.

1. Fédération Française des Diabétiques – "Traitement innovant du diabète : à la recherche du pancréas artificiel"
2. Developed by the company Diabeloop
3. Pierre-Yves Benhamou, Sylvia Franc, Yves Reznik, Charles Thivolet, et al. (Diabeloop WP7 investigators). Closed-loop insulin delivery in adults with type 1 diabetes in real-life conditions: a 12-week multicentre, open-label randomised controlled crossover trial – Lancet Digital Health 2019; 1: e17–25