Can digital twins be built for patients with type 1 diabetes (T1D) to define clinical phenotypes and simulate each person’s risk for diabetes-related complications? That is the question researchers in Luxembourg and France are trying to answer through analysis of data from 2187 participants in the SFDT1 cohort.

Guy Fagherazzi, PhD, director of the Department of Precision Health and head of the Deep Digital Phenotyping Research Unit at the Luxembourg Institute of Health in Strassen, Luxembourg, and the study’s principal investigator, told Medscape’s French edition that the system is being developed as a clinical decision-support tool to complement conventional risk scores and help guide not only prevention strategies but also personalized treatments and therapeutic education.

“This very first model, which we presented at this European Association for the Study of Diabetes Annual Meeting (EASD) 2026, allowed us to describe the profiles of people living with T1D and to assess their risks of complications. The goal is to enable personalized risk prediction. In the longer term, we could also use digital twins for therapeutic optimization,” he said.

Patient Data Collection

The French SFDT1 cohort, which began in 2021 and has recruited nearly 7000 participants to date, is one of the largest T1D cohorts in the world. “That’s why it’s particularly interesting. This data comes from regular clinical visits at the hospital, which we’ve supplemented with information provided directly by the participants through questionnaires,” he said.

The population analyzed in this study consisted of an equal number of men and women, with a mean age of 40 years, who had been living with diabetes for at least 20 years. “In this study, we focused on adult participants because the data collected in pediatric settings are somewhat different. But we will certainly extend this approach to younger participants in the future.”

Modeling Using AI

Using a machine learning algorithm called discriminative dimensionality reduction via learning a tree, the researchers applied 25 variables — sociodemographic, clinical, and biochemical variables and those derived from continuous glucose monitoring — and constructed a phenotypic structure that resembles a star or a tree, reflecting the heterogeneity of patients with T1D.

The researchers identified seven phenotypes among patients with T1D, with profiles ranging from well-controlled disease or very low risk to profiles with a very high risk for cardiovascular or renal complications. This risk is characterized, among other factors, by poor glycemic control, being overweight, social vulnerability, advanced microvascular and macrovascular complications, age at diagnosis, A1c, time spent within target ranges, lipid profile, smoking, and high blood pressure.

Thus, patients with T1D who have similar A1c levels but different complication profiles occupy distinct regions in this seven-branch tree and will therefore likely experience a disease course that progresses differently. A web interface allows users to view real-time projections and simulations of clinical profiles. Specifically, each person is mapped onto this model and can navigate the tree based on changes in characteristics, such as glycemic parameters.

Model Validation and Limitations

One of the inherent challenges of AI is validating model performance. In this project, in addition to the first 2187 participants used to train the model, the researchers selected an additional 600 individuals whose data had not been used.

The researchers applied the model to these 600 participants and measured the distance — that is, the error rate — between the model’s projection and the patient’s actual position on the modeled tree. “We were able to confirm that the validation was very good. The model is very stable when applied to new patients it has never seen before,” Fagherazzi said.

To refine the model, however, more detailed data will be needed, particularly regarding lifestyle. Indeed, the researchers would like to go further in characterizing patients with better information on diet, sleep, and other factors as well as genetic data.

What about randomness in the model? “There is always some randomness in models, and by definition, it cannot be fully explained. So there is always some variability in the data,” Fagherazzi said.

When asked about data security, Fagherazzi explained that the data are “pseudonymized, stored, and hosted in France on the secure servers of a contract research organization. This is an entirely academic study, approved by France’s data protection authority.”

A New Tool for Tomorrow’s Healthcare Professionals?

Eventually, healthcare providers will be able to enter the parameters of the person in front of them and see where that person falls on the branches of the tree, that is, which broad risk category they belong to. The goal is not to replace existing risk-scoring models but to use this digital twin as a simulation tool for a specific individual, with virtual tests.

Thus, a doctor could predict changes in renal, neurologic, and cardiovascular risk by modifying certain parameters: What would happen if the patient could lose x kg? What if we lowered their A1c by x% or even prescribed a specific treatment?

Fagherazzi also emphasized the tool’s educational potential, saying that they will be able to tell patients: “Look, on this graph, if you lower your cholesterol by this much, your risk will drop by this much.” Visually, patients will be able to see their profile change.

But he also seeks to reassure people about the role of AI in medical decision-making: Like any AI model, we must pay attention to its performance and validity. It’s not perfect; some people might be misclassified or mispredicted. So we must always keep the human in the loop because the tool is not intended to make decisions in place of the healthcare professional.

This story was translated from Medscape’s French edition.