Univariate and multivariate time series blood glucose prediction with LSTM deep learning model

• Main Authors: Mohamed Nordin, Muhammad Shahmi, Mahmud, Farhanahani
• Format: Conference or Workshop Item
• Language: English
• Published: 2024
• Subjects: QA71-90 Instruments and machines
• Online Access: http://eprints.uthm.edu.my/12087/1/P16855_de5069adb19c0e9bbe34a3a653fbd10c.pdf%205.pdf; http://eprints.uthm.edu.my/12087/; https://doi.org/10.30880/eeee.2024.05.01.035

Type-1 diabetes is a chronic autoimmune condition impacting insulin production, a crucial hormone for regulating blood sugar levels. Caused by the immune system mistakenly attacking insulin-producing beta cells in the pancreas, this leads to insulin deficiency and elevated blood sugar levels. Globally, approximately 9 million people, representing 0.1% of the population, grapple with type-1 diabetes. With technological advancements, a need arises for the blood glucose prediction model to help monitor and manage type-1 diabetes patients. Deep learning, particularly Long Short-Term Memory (LSTM) networks, proves its ability to grasp long-term dependencies in sequential data however, creating an accurate model for a time-series blood glucose prediction is a complex challenge that requires further research and exploration in the modeling. Thus, leveraging the Cobelli model to simulate blood glucose data in type-1 diabetes patients, the primary goal is to utilize an LSTM network for better prediction of glucose levels. Ten datasets containing information on blood glucose levels, insulin, and meal intake are employed to train both univariate and multivariate models. The univariate model relies solely on glucose data, while the multivariate model integrates insulin and meal intake variables. Two prediction horizons (5- and 10-minute) are utilized to assess and compare model performance. Performance evaluation includes regression analysis metrics of Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and MAE. From the result, it is found that the multivariate model has shown a better prediction performance compared to the univariate model, with the best mean error scores of 0.8777 [mg/dl] for the MAE, 0.958024 [mg/dl] for the RMSE and 1.9875 [mg/dl] for the MSE with the 5-minute prediction horizon outperformed the 10-minute prediction horizon. Based on the findings, a better understanding of designing a high-performance LSTM deep learning model for blood glucose prediction has been achieved, which could promote better diabetes treatment