Researchers have launched the MELISSA trial to determine whether an artificial intelligence–driven insulin-dosing platform can improve time in range among people with type 1 diabetes treated with multiple daily insulin injections. The study’s protocol was published in Diabetes, Obesity and Metabolism.  

The MELISSA trial is a 22-week, prospective, randomized, open-label, blinded-endpoint study being conducted across four diabetes centers in the Netherlands, Denmark, and Greece. Researchers plan to enroll 278 individuals with type 1 diabetes and an exploratory cohort of 50 patients with type 2 diabetes who have been using multiple daily insulin injections for at least 1 year.  

After a 6-week run-in phase, participants will be randomized 1:1 to receive either the MELISSA system or continue usual care.  

The primary outcome is the difference in time spent in the target glucose range—3.9 to 10 mmol/L (70 to 180 mg/dL)—between the trial arms from weeks 14 to 22, as measured by blinded continuous glucose monitoring.  

The MELISSA system consists of two artificial intelligence (AI) features: the Adaptive Basal-Bolus Advisor (ABBA), which an automated dietary assessment system, goFOOD, can support. ABBA uses reinforcement learning algorithms to generate customized recommendations for basal and mealtime insulin dosing based on glucose values, insulin-on-board, carbohydrate intake, and prior glycemic responses. goFOOD estimates carbohydrate content from smartphone images of meals using machine learning–based food recognition and three-dimensional volume estimation.  

During the intervention phase, individuals assigned to MELISSA will receive adaptive insulin recommendations through ABBA and optional carbohydrate estimation using goFOOD. The control group will continue with usual care while using an inactive version of the MELISSA application for logging glucose, insulin, and carbohydrate data.  

Secondary outcomes include consensus and non-consensus continuous glucose monitoring metrics, such as time above range, time below range, time in tight range, and hypoglycemic events. Researchers will also assess treatment satisfaction, diabetes numeracy, fear and awareness of hypoglycemia, diabetes distress, quality of life, and perceptions of AI-based insulin advisors. Safety outcomes will include severe hypoglycemia requiring assistance, diabetic ketoacidosis, hospitalizations, frequency of rescue recommendations, and acceptance or modification of insulin dose suggestions.  

The study is powered to detect a 5% difference in mean time in range between groups, a threshold researchers described as clinically meaningful.  

Outcomes will be analyzed using linear mixed-effects models under an intention-to-treat framework, with additional sensitivity analyses examining missing data and the independent contribution of automated carbohydrate estimation.  

Researchers will also collect biometric data from wearable devices, laboratory biomarkers including hemoglobin A1c, lipids, inflammatory markers, and advanced glycation end products, and conduct a health technology assessment to evaluate cost-effectiveness.  

Trial completion is expected in 2027.  Researchers can follow the trial for updates at www.melissa-diabetes.eu.  

The study is being funded by the European Union Horizon Europe Research and Innovation Actions program, the Swiss Confederation State Secretariat for Education, Research and Innovation, and Breakthrough T1D. Dexcom International will provide continuous glucose monitoring devices, and Novo Nordisk is supplying study insulins and smart pens. The funders had no role in the study design and are not expected to participate in trial conduct, analysis, interpretation, or publication decisions.  

Researcher Ulrik Pedersen-Bjergaard, MD, reported speaker fees from Sanofi and Novo Nordisk and advisory board participation for Sanofi, Novo Nordisk, and Vertex. The remaining researchers reported no conflicts of interest. 

Expert Insight 

AACE Endocrine AI invited researcher Lisa den Brok with the CARIM School for Cardiovascular Disease at Maastricht University in Maastricht, the Netherlands, to expand on the trial’s goals and objectives. 

Why does this trial matter? 

Over the last decade, innovative diabetes technologies, such as hybrid closed-loop systems, have substantially improved diabetes self-management. Nevertheless, achieving glycemic targets remains an ongoing challenge for many individuals with diabetes, particularly those requiring insulin therapy. Calculating the appropriate dose of insulin requires adjustment based on multiple dynamic and personal parameters, including carbohydrate intake, physical activity, illness, and stress. This complex skill is a time-consuming and error-prone task, with limited health literacy and fear of hypoglycemia posing potential obstacles. 

Artificial intelligence has promising potential to further support people with diabetes in making this complex decision. This study is among the first randomized trials evaluating an AI-based insulin support system in the context of obtaining CE certification (indicating that a product complies with European Union health, safety, and environmental protection standards). It aims to determine whether such an integrated AI-based approach can improve glycemic control while reducing the cognitive burden of daily diabetes self-management, addressing a key unmet need in current diabetes care.  

Is there any preliminary data that has stood out? 

One early observation was that improvements in glycemic outcomes were already seen in the intervention group within the first few weeks of the trial. Although these findings are preliminary, they are consistent with our results from a smaller pilot study conducted in Switzerland and suggest that AI-based support systems may help people with diabetes achieve their recommended treatment targets.  

In addition, the trial has highlighted the importance of conducting a randomized clinical trial with a longer study duration for CE-marking. In silico and/or pilot studies do not always capture certain issues within a technical system that only become visible in real-life settings. These new insights are currently integrated into the algorithm to improve the next version of the MELISSA system.  

How might the findings influence endocrinology clinical practice? 

If the preliminary results are confirmed, they could have several important implications for endocrinology clinical practice.  

First, AI-based insulin decision support systems like MELISSA may become a valuable addition to standard diabetes care by helping people with diabetes accurately adjust their insulin doses in daily life. By continuously adapting to an individual’s behavior and physiology, the MELISSA system could improve glycemic control, increase quality of life, and lower the risk of diabetes-related complications.  

Second, the MELISSA system has the potential to reduce healthcare burden and associated economic costs. By enabling more precise, individualized treatment, healthcare professionals may be better able to predict patient responses, potentially reducing the need for frequent follow-ups. If successfully implemented and CE-certified, the MELISSA system also could be integrated into routine care as a complement to existing diabetes technology, such as sensors and insulin pumps.  

Finally, the MELISSA system is designed to be accessible to individuals from diverse socioeconomic backgrounds, as it can integrate data from both sensors and glucose meters. In addition, it will be available for people on multiple daily injections, which remain the most widely used treatment modality worldwide.