Evaluation of a Mobile AI-powered Decision Support System for Insulin Dosing and Glucose Prediction in Type 1 Diabetes: The glUCModel Clinical Trial Protocol

Overview

The goal of this clinical trial is to evaluate the safety and efficacy of integrating predictive models into insulin therapy management via the user-centered glUCModel mobile app in People with Type 1 Diabetes Mellitus following Multiple Insulin Dosing therapy. Participants will be aged 18 to 65 years. The main questions it aims to answer are: Does using the app improve glycaemic control, as measured by time in range? Does using the app reduce the number of episodes of hyperglycaemia and hypoglycaemia? Are the app's design and functionality adequate? The study will comprise four phases:ses}): * Screening phase: Informed consent, collection of sociodemographic and clinical data, and baseline Pittsburg, IFIS, and DTSQ questionnaires. * Run-in phase: 2 weeks of standard care with CGM. Data will be used to generate personalized predictive models in the intervention group. * Active treatment phase: Participants continue MDI therapy. The intervention group will additionally use the glUCModel mobile app. CGM data from the final 2 weeks will be analyzed. * Evaluation and analysis phase: Participants will complete the uMARS, Pittsburgh, and DTSQ questionnaires. Statistical analysis and correlations among outcomes will be processed.

Diabetes mellitus is a chronic, metabolic disorder characterized by impaired regulation of blood glucose, affecting more than 400 million people worldwide. Insulin, a hormone produced by the pancreas, facilitates the uptake of glucose into cells for energy production. In diabetes, either insufficient insulin is produced or the body cannot use it effectively, leading to persistent hyperglycemia. Over time, uncontrolled glucose levels can result in serious complications, including cardiovascular disease, neuropathy, retinopathy, and nephropathy. Effective management is therefore essential to prevent both acute and long-term adverse outcomes. Two main forms of diabetes can be distinguished. Type 1 diabetes mellitus (T1DM) is an autoimmune condition in which pancreatic β-cells are destroyed, resulting in absolute insulin deficiency. It accounts for approximately 10% of all cases. Individuals with T1DM require lifelong insulin replacement therapy, typically delivered as multiple daily injections (MDI) or via an insulin pump. In contrast, type 2 diabetes mellitus (T2DM), the more prevalent form, is characterized primarily by insulin resistance. While insulin production is preserved in early stages, progressive dysfunction may ultimately necessitate pharmacological therapy, including insulin. Lifestyle interventions such as healthy diet and physical activity can delay or prevent T2DM onset and progression. For individuals with diabetes, day-to-day self-management requires frequent glucose monitoring and insulin dose adjustments that must take into account meals, physical activity, stress, illness, and other factors. Capillary glucose meters and, more recently, continuous glucose monitoring systems (CGMs) have greatly improved access to real-time glucose data. However, interpreting these data and deciding on corrective actions remains challenging, and errors in insulin dosing can lead to hypoglycemia or persistent hyperglycemia. Both acute complications and the constant decision-making load contribute to reduced quality of life and treatment fatigue. To support patients in these complex tasks, predictive models of glucose dynamics have been extensively investigated. Accurate prediction could enable early warnings of hypo- or hyperglycemia and assist in optimizing insulin therapy. The ultimate vision is the development of a fully automated ''artificial pancreas'' combining glucose sensing, insulin delivery, and robust prediction algorithms. Various machine learning (ML) approaches have been explored for glucose forecasting, including Genetic Programming , K-Nearest Neighbours , Grammatical Evolution, and, most prominently, Neural Networks. Among neural architectures, Long Short-Term Memory (LSTM) and other recurrent models have demonstrated strong performance for time-series data such as CGM traces, although convolutional and multilayer perceptron (MLP) networks have also been applied. Despite encouraging results, challenges remain in ensuring accuracy, robustness, and real-world usability across diverse patient populations. Managing T1DM, particularly in patients using MDI, continues to pose a major challenge. While CGM and insulin pumps have improved outcomes, decisions about insulin dosing still depend heavily on patient intuition and experience, leaving room for error and variability. There is therefore a clear need for decision-support tools that combine predictive analytics with personalized recommendations to enhance safety, autonomy, and treatment adherence. The glUCModel mobile application was developed to address this need. Since its early versions, it integrates proprietary, patented artificial intelligence models to provide real-time insulin recommendations, short-term glucose forecasts, and predictive alerts for hypo- and hyperglycemia. With a forecast horizon of up to two hours, the system aims to reduce glycemic variability and support timely corrective actions. This protocol describes a randomized, open-label clinical study to evaluate the efficacy and safety of the glUCModel application in patients with T1DM using MDI therapy. The primary objective is to assess improvement in short-term glycemic control, measured by the percentage of time spent in target range (70-180 mg/dL). Secondary objectives include reductions in glycemic excursions, improved treatment satisfaction, and evaluation of usability and adherence in a real-world setting.