Exploring the Potential of Deep Learning Models Integrating Transformer and LSTM in Predicting Blood Glucose Levels for T1D Patients

Abstract Type 1 diabetes (T1D) is an increasingly common chronic disease, and its incidence continues to rise globally, posing huge challenges to the public health system. At the same time, due to the need for long-term strict diet control and blood sugar monitoring, the management process of diabetes is relatively complicated, which brings a lot of inconvenience to patients' daily life. Predicting blood glucose concentration (BGC) in T1D patients can help avoid abnormal blood glucose events and improve blood glucose management in T1D patients. However, accurate blood glucose prediction remains challenging because blood glucose levels are affected by multiple factors such as insulin injections, diet, and exercise. In this study a combined deep learning (DL) model of Transformer and long short-term memory network (LSTM) for blood glucose prediction in T1D patients is proposed, which fully integrates the advantages of Transformer and LSTM. For the proposed method Transformer captures sequence relationships globally, while LSTM focuses on local and long-term dependencies, thereby improving the overall performance of the model. In this study the clinical data set of 13 T1D patients for more than 4 weeks are provided by the First Renmin Hospital of Yunnan Province, and the 360-day blood glucose data set of 10 adult T1D patients generated by the UVa/Padova simulator. The prediction performance and clinical evaluation were performed using root mean square error (RMSE), mean absolute error (MAE), and Clark error grid analysis (EGA), respectively. The performance of the model on clinical data sets is 30-min prediction horizon (PH) (RMSE=10.157, MAE=6.377), 60-min PH (RMSE=10.645, MAE=6.417), 90-min PH (RMSE=13.537, MAE=7.283), 120-min PH (RMSE= 13.986, MAE= 6.986). The performance of the model on the simulated data set is 15-min PH (RMSE=1.793, MAE=1.376), 30-min PH (RMSE=2.049, MAE=1.311), 60-min PH (RMSE=3.477, MAE=1.668). The units of RMSE and MAE are mg/dl. In addition, Clark Grid Analysis (EGA) analysis showed that more than 96% of BGC predictions remained within the clinical safety zone during the PH period of up to 120 minutes in the clinical data set, validating its clinical feasibility.