Prediction of Blood Glucose Levels in Patients with Type 1 Diabetes via LSTM Neural Networks

Diabetes is one of the most prevalent diseases of the 21st century, with more than 500 million people affected. Having tools to estimate blood glucose levels is critical for these patients in their management of the disease. In this work, we present a comparison of three neural network architectures based on long short-term memory (LSTM). Their predictive ability has been evaluated against a longitudinal dataset with continuous glucose level measurements of patients with type 1 diabetes. All models, trained for different prediction horizons of 30, 60, 90 and 180 min, have generally yielded good prediction results. These results are further validated using clinical standards resulting in more than 95% of accurate blood glucose level predictions, mostly leading to correct treatments.