Fractional-Order Neural Networks for Data-Driven Model Predictive Control of an Artificial Pancreas

Accurate prediction of blood glucose trajectories is essential for safe automated insulin delivery and clinical decision support in type 1 diabetes. Yet existing data-driven methods based on historical patient data perform poorly in predicting long-term outcomes. This work presents a fractional-order neural network (FONN) framework for the identification and control of blood glucose dynamics in patients affected by type~1 diabetes. The proposed approach integrates Grünwald–Letnikov operators into the neural network architecture, enabling the inclusion of long-term memory effects into the learning process. In open-loop system identification, the FONN is benchmarked against its integer-order ablation (IONN), and a Long Short-Term Memory (LSTM) model. On a widely used in-silico simulator, LSTM attains the best short-horizon forecasts, whereas FONN yields higher accuracy as the prediction horizon increases. Closed-loop evaluations in the context of data-driven model predictive control confirm the advantages of the FONN compared to a linear model-based formulation and the LSTM approach itself, with improved performance in terms of time in normoglycemia and euglycemia.