Artificial Intelligence Models for the Management of Type 1 Diabetes

Type 1 Diabetes mellitus (T1D) is a chronic metabolic disease due to which the pancreas is not able to produce an adequate amount of insulin, resulting in an increased blood glucose concentration. If not treated properly, it can lead to short- and long-term complications requiring emergency care and life-threatening conditions. The advent of Continuous Glucose Monitoring (CGM) sensors has considerably improved the management of T1D, as it allows people suffering from this disease to monitor their glycemic levels for 24 hours a day. These sensors are usually coupled with an insulin pump, a device able to continuously provide small amounts of subcutaneous insulin and larger amounts at the patient’s request. Since the final decision on glycemic control is taken by the patient, who is a part of the control loop, such a device is defined as a hybrid closed-loop artificial pancreas. In the last decade, CGM data have been utilized together with Artificial Intelligence (AI) and time-series techniques with the aim of improving T1D management and increasing the quality of life of people with T1D. In this frame, regression is by far the most widely investigated task. In practice, CGM and other features such as injected insulin are given as input to a predictive model in order to forecast future glycemic levels; in this way, the patients are warned in advance of what their blood glucose level is going to be in the next future and are thus able to take the appropriate countermeasures if the glycemia is predicted to exit the target range. A different approach resorts to classification, in which the AI model is trained to predict whether or not the patient is going to experience an adverse event, without predicting the exact value of the future glycemic level. These studies are relevant because they usually achieve better accuracy with regard to the prediction of hypoglycemic events compared to the regression approach. While regression and classification limit to provide the patient with a decision support based on the prediction of future glycemic levels or events, leaving to the patient the management of the disease, a third approach focuses on the control of glycemia, i.e., decides what is the optimal amount of medication that must be provided in order to maintain the blood glucose level within the target range. This manuscript aims to provide significant and several contributions in the field of the application of AI methodologies to T1D management. With regard to the regression task, a novel neural network is presented for the forecasting on adult patients during daily-life activity, and the comparison of different learning techniques is performed on data of patients during sports; the optimal amount of data for training an AI algorithm for the application on an edge-computing device is investigated; an edge-computing application is developed for the forecasting of glycemic levels of pediatric patients. With regard to the classification task, a layered meta-learning approach is presented for the prediction of hypoglycemic and hyperglycemic events of adult patients during daily-life activity and during sports, and the system is implemented on an edge-computing device. With regard to the control task, a new glycemic closed-loop control based on Dyna-Q is presented that does not necessitate information on carbohydrates, thus not requiring any human intervention and providing a fully closed-loop control.