Predicting Blood Glucose Levels with EMD and LSTM Based CGM Data

Abstract

Blood glucose monitoring is essential for diabetes management. Applying deep learning technique for blood glucose monitoring is promising, given its success in a range of healthcare and medical tasks. In this paper, we proposed a method that combines Empirical Mode Decomposition (EMD) with Long-Short Term Memory (LSTM) to achieve good experimental results in predicting patient blood glucose. We used patients' real blood glucose levels time series data to train the method proposed in this paper and to predict blood glucose for 30 minutes to 120 minutes. First, we use only blood glucose readings and timestamps in the dataset. Meanwhile, we used ADF to verify the non-stationarity of blood glucose time series. Then, we use EMD to decompose the blood glucose time series and use LSTM to train the decomposed time series to obtain a blood glucose prediction model. Finally, Mean Absolute Error (MAE) and root mean squared error (RMSE) were used to evaluate the experimental results. On the test dataset, the mean values of the MAE and RMSE are 0.4458mmol/L and 1.08mmol/L for 30mins, 0.87 and 1.27 mmol/L for 60mins, 0.85mmol/L and 1.36 mmol/L for 120mins, respectively. Experimental results show that the EMD+LSTM had better predictive performance than the LSTM when blood glucose changed dramatically. Meanwhile, it is still challenging to reach a high accuracy of predicting the long-term blood glucose.