Development of attention-based LSTM models for the prediction of nocturnal hypoglycemia in patients with Type 1 Diabetes

Abstract

Diabetes Mellitus (DM) is a chronic condition with a rising global prevalence and severe complications. The International Diabetes Federation projects that the number of individuals with diabetes will reach 643 million by 2030. To enhance glycemic control and mitigate the risk of serious physical and emotional complications related to hypoglycemia, this thesis presents the design, development, and evaluation of an interpretable model for predicting the risk of nocturnal hypoglycemic episodes in individuals with Type 1 Diabetes (T1DM). The proposed model employs a hybrid approach, integrating compartmental models with machine learning techniques. The OhioT1DM dataset, which includes real data from the eight-week monitoring period of 12 patients, was utilized for both development and evaluation purposes. Input data for the model consisted of glucose measurements, insulin doses, and meal information from the previous 24 hours. Mathematical models for simulating (i) the physiological mechanisms of insulin absorption from the subcutaneous tissue into the bloodstream, (ii) the activation of the insulin signaling pathway, and (iii) the absorption of glucose from the intestine were combined with Long Short-Term Memory Neural Networks (LSTMs). A custom attention layer was integrated to enhance the model’s performance and provide insights into the model’s reasoning behind its predictions. The model was assessed in terms of its ability to correctly predict nocturnal hypoglycemic events within a twelve-hour prediction window. Moreover, the Monte Carlo Dropout method was applied to quantify the uncertainty of the model's predictions. The model was also evaluated on an external dataset from the ten-day monitoring period of 12 T1DM patients, which was granted from the Diabetes Center, First Department of Pediatrics, P. & A. Kyriakou Children’s Hospital, Athens, within the framework of the SMARTDIAB project.