Accurate forecast of blood glucose levels could predict the onset of hypo- or hyperglycemic events and provide actionable information to patients, thereby improving glycemic variability and time in range, two important metrics for managing diabetes and assessing clinical outcomes. Current commercial continuous glucose monitors (CGMs) offer glucose trend prediction, but these predictions are often inaccurate. We hypothesize that the limited performance and utility of these predictive algorithms is likely due to using only glucose as the explanatory variable. By accounting for additional relevant features, we aim to build an algorithm that accurately predicts glucose levels with a prediction horizon of 1 hour. We conducted a pilot study in which blood glucose levels, meal and insulin logs, and activity-related parameters, like heart rate and steps, were collected from 5 individuals over a period of 2 weeks. This data was used to develop algorithms for glucose level prediction based on a number of models, including: long short-term memory (LSTM) recurrent neural network (RNN), Gaussian Process (GP), Lag Method and Clockwork RNN. The predictions using the LSTM model yielded the best performance. With the inclusion of additional features, the RMSE for a 1 hour prediction horizon using the LSTM RNN was 8.71 ± 3.98 mg/dL. However, work must still be done to improve performance and determine the importance of each feature.
D. Colburn: None. V. Behnam Roudsari: None. R. Gandica: None. S. Sia: None.