CGM-Based Blood Glucose Prediction Model With LSTM Encoder–Decoder Architecture

Accurate prediction of blood glucose levels is crucial for automated treatment in diabetic patients. This study proposes a blood glucose prediction model based on an improved attention mechanism within a long short-term memory (LSTM) encoder-decoder (Att-E-D) architecture to enhance blood glucose prediction performance significantly. Compared to traditional encoder-decoder (E-D) architectures, the core improvement of this study lies in introducing an attention mechanism combined with a dynamic time warping (DTW) similar sequence search algorithm. Specifically, during the decoding phase of the Att-E-D architecture, the DTW-based similar sequence search algorithm is first utilized to retrieve the n most matching similar sequences from historical blood glucose data for the target prediction sequence. Then, at each time step, different attention weights are assigned to the encoded information. This enables the model to selectively focus on historical feature information with higher similarity to the current step, effectively enhancing neuron memory and reducing error propagation. Finally, after processing through a fully connected layer, the model outputs a prediction sequence for blood glucose trends over a future period. To validate the generalization ability of the Att-E-D model, comparative experiments were conducted using datasets from two different continuous glucose monitoring (CGM) sensors. The results demonstrate that the Att-E-D model significantly outperforms support vector regression (SVR), LSTM, gated recurrent unit (GRU), bidirectional LSTM (Bi-LSTM), Bi-GRU, and the basic E-D model in prediction accuracy, achieving the highest ${R}^{{2}}$ values of 0.952 and 0.972 on the two datasets, respectively, proving its superior ability to capture the long-term dependencies in blood glucose data.