In 2018, the US Food and Drug Administration (FDA) for the first time approved an autonomous artificial intelligence (AI) diagnostic system — one to detect diabetic retinopathy. Today, the FDA and the European Medicines Agency (EMA) have approved over 500 AI systems in various medical fields with the hope that AI could optimise treatments in unprecedented ways. Provided it is implemented carefully.

“Through a Delphi procedure, we have identified 14 recommendations that we felt should be considered when implementing AI-CDSS for the management of diabetes,” explains Bart Torbeyns, co-author from the European Diabetes Forum (EUDF), based in Brussels, Belgium.

Why diabetes is a key testing ground

The EUDF set up a working group with members of not-for-profit diabetes associations and federations, diabetes physicians, medical technology experts, industry stakeholders, and individuals living with diabetes. Their consensus focuses on finding the sweet spot between optimising clinical outcomes and using AI systems to tackle treatment inertia – with the aim of stimulating debate within the community. Given the rich volume of available data and clearly defined treatment targets, diabetes care is a particularly promising environment for harnessing the benefits of AI applications, they argue.

The role of clinicians in shaping AI

One major emphasis of the consensus is on the role of clinicians for the implementation of AI-CDSS, as pointed out by working group member and last-author of the consensus, Stefano Del Prato, who is Professor of Medicine at Sant’Anna School of Advanced Studies in Pisa, Italy.

As Del Prato puts it: “The functionality and reliability of these systems largely depend on the data they are educated on, and we are the ones who should ensure top quality clinical data,” he points out, calling on clinicians to stay up to date with AI. “Engage in continuous professional development courses after your qualification and ensure your periodic recertification in this fast-moving discipline. Understanding the principles and practice of AI-CDSS is a core competency for medical professionals.”

How AI supports diabetes care

In which ways could AI actually benefit diabetes care? Examples support glucose control self-management, making predictions about complications, and managing medications in cases of polypharmacy. AI systems trained on clinical recommendations could also extract personalised treatment pathways using patient-specific data.

The working group argues that such applications could benefit people with diabetes and healthcare professionals (HCPs), while reducing healthcare costs.

“There is a fast-growing number of people with diabetes and an ever-decreasing number of HCPs tasked with their care,” says Torbeyns from EUDF. “Ideally, the use of AI-CDSS leads to improved workflows and reduced administrative burdens and empowers HCPs to focus on human factors in the delivery of care.”

14 recommendations for five goals

The working group defined 14 recommendations for five overarching goals, including covering the unmet needs of people with diabetes, training medical professionals, establishing regulatory processes, and setting data standards for development and clinical trials.

“Among these 14 recommendations, I believe the one requiring the greatest urgency and consideration is the third, calling for the acknowledgement of the person-centred objectives of equity, personalisation, integration with healthcare services and evaluation,” says Torbeyns.

This includes ensuring that AI systems are accessible to all people with diabetes and easy to use. Individuals should be able to adjust systems to their specific needs and personal circumstances. Also, AI-CDSS should facilitate close interaction between healthcare teams and people with diabetes to optimise care, and the systems themselves must undergo rigorous evaluation (see Figure 1).

Torbeyns emphasises: “All these elements are key in reducing the burden of the disease for people with diabetes, while supporting diabetes personalised delivery of care, as well as rationalising and optimising the use of resources.”

Safety, trust and regulation

Despite the potential of AI, the consensus acknowledges that significant challenges remain, such as interoperability between technologies, supporting primary care teams in adopting the systems, and giving people equal and easy access to AI-CDSS. From a clinical point of view, one challenge is to judge whether an AI-CDSS is ‘good enough’ to trust in routine care, as pointed out by Stefano Del Prato.

“The biggest issue we see is the phenomenon of ‘hallucination’, whereby an AI model produces factually incorrect or nonsensical responses, which may represent a risk particularly if not associated with educated and critical use of the AI tools,” he explains.

Therefore, Del Prato advocates rigorous validation of AI-CDSS within real-world clinical workflows prior to widespread deployment.

„Along with that, as highlighted in our recommendations, objective criteria must be developed to support risk categorisation assignment with the data on which the European market authorisation is based available for public scrutiny,” he argues.

Looking ahead: predictive AI

As AI is rapidly transforming many professions, the discussion on AI in diabetes care has only just begun.

“So far, the accomplishments we see in diabetes are the use of routinely collected data from AI-CDSS for managing and optimising the treatment of people with diabetes, as well as the use of AI-assisted systems for the screening of diabetic retinopathy,” says Torbeyns. “Yet these are just specific applications of AI while from a public health perspective a wide application of predictive AI based on Large Language Models may improve the management of chronic conditions such as diabetes.”

AI is used to carry out tasks that typically require human intelligence, such as pattern recognition. Generative AI, a subset of AI best known from large language models (LLMs), takes this further by producing meaningful content – potentially acting as a ‘clinical predictive engine’.

“Generative AI should be adapted and made appropriate for implementation in primary care settings to support training for primary care physicians, increase their knowledge and skills, streamline clinical tasks and improve clinical decision-making. This can include risk assessment and diagnosis, disease management, prediction of disease progression, individualising intervention and optimising early measures of treatment response,” outlines Torbeyns.

Above all, however, clinicians must play an active role in developing AI-CDSS — and take ownership of this responsibility.

Citing Michael Howell, Google Health’s chief clinical officer,Del Prato emphasises: “We talk about how AI is going to change a lot of things and it should happen with clinicians, not to clinicians.”

Key Points:

• AI-driven clinical decision support systems (AI-CDSS) could improve treatments and clinical outcomes, while reducing resource needs and administrative burden.
• A new Delphi-based consensus roadmap provides guidance for the effective and safe implementation of AI-CDSS in diabetes care.
• 14 recommendations were identified, clustered into five goals to enable patient-centred, robust and safe AI-CDSS.
• Safe implementation requires rigorous real-world validation, clear risk categorisation and high-quality clinical data.

To read this paper visit: Bajramagic M, Battelino T, Cos X, et al. Artificial intelligence-driven clinical decision support systems to assist healthcare professionals and people with diabetes in Europe at the point of care: a Delphi-based consensus roadmap. Diabetologia. 2026;69(2):259-273. 

Author: Hanna Gabriel, BA MSc. Any opinions expressed in this article are the responsibility of EASD e-Learning.