Personalized LSTM-Based Alarm Systems for Hypoglycemia Prevention in an Intraperitoneal Artificial Pancreas

Prevention of hypoglycemia is a key aspect for efficient management of Type I diabetes. Alarm systems (ASs) are very useful to alert the patient in advance in case of hypo-glycemia, allowing early intervention to avoid or mitigate the potential critical situation. Model-based ASs use patient models to predict the future glucose concentration and trigger alarms. In recent years neural networks, in particular Personalized Long Short-Term Memory Networks (PLSTMs) have shown very promising performances in glucose prediction. In this work, PLSTM-based AS for the prevention of hypoglycemia for an Artificial Pancreas is proposed. Preliminary results on a subgroup of 71 patients show that this system is able to predict almost all the potentially critical events (median TPR = 100%) with a precision of 57%. These promising techniques are under study to include also the remaining 29 problematic patients.