Comparative Analysis of LSTM and GRU Neural Networks in Predicting Hyperglycemic and Hypoglycemic Events

A Time Series Approach for Diabetes Patient Monitoring

Abstract

The research aims to compare the effectiveness of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks in predicting hyperglycemic and hypoglycemic events in diabetes patients using wearable sensor data, to identify the more effective model for near-term blood glucose forecasting. Effective prediction of blood glucose levels is crucial for Type 1 diabetes management, as extreme glucose levels can cause fatal hyperglycemic and hypoglycemic incidents. Choosing between LSTM and GRU models can significantly impact predictive accuracy and patient outcomes.

The OpenD1NAMO public dataset was used as a training set to compare the efficacies of LSTM and GRU models in predicting near-term glucose levels. Artificial features such as lagged features and rolling averaged lagged features were created to facilitate glucose forecasting comparisons. Various permutations of hyperparameters were then compared to find the most effective set for the provided sensor data. GRU and LSTM models trained on a portion of the preprocessed dataset were compiled and the average of the minimum MSEs per epoch were compared.

On average, the LSTM model was 53.5% more accurate than the GRU model for the dataset provided, forecasting blood glucose values with a 15-minute interval. The minimum average MSE of the LSTM model was 2.043 mmol2/L2 while the minimum average MSE of the GRU model was 3.137 mmol2/L2.

Given the models’ performances on the dataset, it suggests that LSTM models will likely be more accurate than GRU models when predicting near-term blood glucose levels using live wearable sensor data.