The project aims to develop and validate a digital twin of glucose dynamics in people with type 2 diabetes (T2D), i.e., a subject-specific model able to reproduce the individual glucose response to meals and therapy, learned from real-world data by means of neural ordinary differential equations (Neural ODEs).

Digital twins of glucose–insulin dynamics are well established in type 1 diabetes, where physiological models and simulators are routinely used for in silico testing of therapies. In T2D, however, the underlying physiology is markedly heterogeneous (variable insulin resistance, residual and declining β-cell function, different combinations of oral agents and insulin), and classical compartmental models are hard to identify from the limited data collected in daily life. Neural ODEs offer a promising alternative: they describe the system in continuous time, as physiological models do, but they can learn the dynamics directly from data with a neural network.

The work will be organized in four consecutive steps, with particular importance devoted to understanding how Neural ODEs work.

1) Literature review on digital twins in diabetes: definitions, existing digital twins and simulators (mainly developed for type 1 diabetes), modeling approaches (physiological, data-driven and hybrid), the data they rely on, and how they are validated.

2) Literature review on physiological models of the glucose–insulin system in T2D: model structure and subsystems (oral glucose absorption, glucose kinetics, insulin secretion and action), how T2D-specific features such as insulin resistance and impaired β-cell function are represented, and how the effect of therapies is described.

3) Development of a Neural ODE on a simple case study: the three-compartment model of the oral glucose absorption subsystem (two stomach compartments and the gut, from meal ingestion to the rate of glucose appearance in plasma). Synthetic data will be generated by simulating this model, and a Neural ODE will be trained to reconstruct its dynamics. Since the ground truth is known, this step will be used to understand how Neural ODEs work in practice: the role of the numerical solver (e.g., Euler, Runge–Kutta, adaptive schemes) and its accuracy/cost trade-off, how gradients are computed (backpropagation through the solver vs. adjoint method), the effect of network size, noise and sampling time, and the ability of the model to generalize to meals not seen during training.

4) Application to the ShanghaiT2DM dataset: the Neural ODE framework will be extended to real-world data to build the digital twin of glucose dynamics in T2D. Inputs will include meal intake, insulin doses and oral hypoglycemic agents; the output will be the CGM signal. Strategies for personalization will be investigated, e.g., a population model fine-tuned on each subject or a shared model conditioned on subject-specific clinical and laboratory variables.

The digital twin will be evaluated in terms of: (i) accuracy in reproducing and forecasting CGM traces (RMSE, MAE, prediction horizons of 30–120 min); (ii) ability to reproduce CGM-derived metrics such as Time in Range, time above/below range and glucose variability; (iii) plausibility of the simulated response to “what-if” scenarios (e.g., different meal size or drug dose). Results will be compared with benchmark approaches, such as linear/ARX models, recurrent neural networks (LSTM) and a physiological minimal model.

The overall goal is to provide a deep understanding of Neural ODEs as a modeling tool for physiological systems and to deliver a validated, data-driven digital twin of T2D that can support in silico evaluation of personalized treatment strategies.