Today, the main use of Artificial Intelligence (AI) in diabetes concerns closed loop algorithms in type 1 diabetes (T1D), to predict the dose of insulin to be delivered. A new wave of even more efficient algorithms is expected: they will eventually make it possible to no longer have to announce meals or physical activity and to learn from patients’ habits (machine learning). Another key use of AI in diabetology concerns the automated detection of ocular complications (diabetic retinopathy) with software capable of analyzing eye funds and detecting and grading diabetic retinopathy: they are used as a diagnostic aid. by an ophthalmologist (deep learning).
THE DAILY: What should we expect in the near future?
Dr FAGHERAZZI: In the more or less near future, we are moving towards the integration of new AI algorithms in different areas of life: enough to imagine, for example, better identifying people at risk of diabetes.
Screening diabetics could therefore be possible using AI, but how?
Our laboratory works a lot on voice as a source of information on people’s health, not only for diabetes, but also for neurodegenerative diseases such as Parkinson’s disease, mental health, cardiorespiratory health, etc. The voice is analyzed and used, either as a screening tool or, in the case of known pathology, for monitoring. We have already shown that people living with diabetes have different voices from the general population (all other things being equal such as age, gender, etc.). It’s not necessarily perceptible to the human ear, but audio signal processing and AI can make a difference.
How do you explain this change in voice?
Chronic hyperglycemia, hypoglycemia, gastric reflux, which is more common in people living with diabetes, chronic fatigue and hydration problems, can play a role. The idea of working on the voice came to us from a few cases of people who had diabetes for more than 15 years and who reported a hoarser voice.
Do you think you will be able to develop a screening tool in the long term?
Yes, our goal is to develop a screening tool to identify people at high risk with 70 to 75% accuracy (but not a diagnostic tool because this tool will not be sensitive enough and not specific enough). Since voice is very easy to collect non-invasively, via a smartphone for example, this tool could be deployed on a large scale. There are more than 530 million people with diabetes in the world, half of whom do not know it: there is therefore a real problem on a global scale of under-screening for diabetes and discovery of the pathology at too early a stage. late, on the occasion of complications.
Can we help this research on voice at our level?
We need people with and without diabetes, over the age of 15, everywhere to donate their voice to our study at colivevoice.org. This includes some health questions and voice recordings. This will be used to develop AI models to detect diabetes and diabetes-related distress (inability to manage your diabetes on a daily basis because it generates too much mental load). It takes approximately 20 minutes for a voice donation. Don’t hesitate to talk to your patients about it!
Can AI still have other applications in screening for complications, for example?
Yes, there are already models capable of predicting the occurrence of a cardiovascular event at 5 or 10 years and work is underway to improve these models with AI. The objective is to be even more efficient in the field of personalized prevention based on calculated risks to intensify their therapy if necessary and better postpone or reduce this cardiovascular risk, which remains the most frequent complication for people living with diabetes. type 1.
Can we imagine tools that will coach people living with diabetes in a very personalized way?
We can indeed imagine connected insulin pens, which over time could make personalized recommendations based on patient profiles, but also smartphone apps that could integrate AI into health and diet recommendations and therefore the choices to be made. in a personalized way. But to be able to individualize advice, you need large validated databases, and that’s what takes time.
What is the main risk linked to the use of these new tools incorporating AI?
The biggest risk is data bias, with training an algorithm on a subgroup of the population thinking that it is generalizable, when not: the tool will then deliver bad advice. Example: take data collected on 50-year-old white men and apply it to a 30-year-old African-American woman! The problem therefore lies in the quality and diversity of data, more than in AI models. The data must perfectly correspond to the target population, in which the algorithm thus developed will be used.
AI algorithms must be stable, robust, explainable: the characteristics of the data used to make the prediction or recommendations must be known to doctors and patients to gain their trust: this is what we call the challenges of explainability.
Is there a risk that malicious hackers will tamper with these tools in exchange for a ransom demand, for example, with risks to the health of people using them?
This is a theoretical risk that cannot be ruled out, but fantasy should not prevent us from moving forward either. Europe adopted the “AI Act”, a regulation on artificial intelligence which, since this year, provides broad recommendations on how to develop tools based on AI while minimizing risks and maximizing benefits. There are different levels of risk classification. Healthcare tools are generally classified as high risk and therefore require developers to have evidence of cybersecurity regarding the use of data, its storage, processes, etc.
What will be the place of generative AI in the future?
New models like ChatGPT which generate content from an enormous amount of data could make it possible to communicate with patients: the latter could ask questions and obtain relevant answers concerning the monitoring of their illness. In our laboratory, for example, we are currently working on a tool for detecting diabetes-related distress, which is insufficiently detected in practice because diabetologists already have a lot to do in consultations. This tool could help prevent and manage this diabetes-related distress.
