Long short-term memory-based model predictive control of blood glucose level for type 1 diabetes mellitus treatment

This paper presents a novel method to control blood glucose levels (BGL) based on predictions made by a long short-term memory (LSTM) network. An initial LSTM model was trained with data from rats with type 1 diabetes mellitus (T1DM) using Open Artificial Pancreas System (OpenAPS). Transfer learning was applied to develop an individualised prediction model based on the initial model. The LSTM model predicted BGL with a root mean squared error (RMSE) of 11.8240 mg/dl. The model was integrated into model predictive control (LSTM-MPC), which optimised insulin injection based on BGL predictions. Evaluated against a neural network-based MPC (NN-MPC) and OpenAPS using different diets and rats, LSTM-MPC outperformed both in control performance. This study demonstrated a closed-loop BGL control system tested with in vivo diabetic rats. The prediction model is re-trainable quickly using small datasets obtained from individual rats, which provides a feasible solution for individualised T1DM treatment.