High-precision blood glucose prediction and hypoglycemia warning based on the LSTM-GRU model

Objective: The performance of blood glucose prediction and hypoglycemia warning based on the LSTM-GRU (Long Short Term Memory - Gated Recurrent Unit) model was evaluated. Methods: The research objects were 100 patients with Diabetes Mellitus (DM) who were chosen from Henan Provincial People’s Hospital. Their continuous blood glucose curves of 72 hours were acquired by a Continuous Glucose Monitoring System (CGMS). The blood glucose levels were predicted based on the LSTM, GRU and LSTM-GRU models, respectively. Analyses of the best predictive model were performed using Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE) and correlation analysis between the prediction blood glucose level and the original blood glucose level acquired by CGMS and Clark Error Grid Analysis (EGA). Repeated-measures analysis of variance (ANOVA) was used to analyze whether the RMSE values of the three models were statistically significant. 60 patients who had experienced hypoglycemia among 100 cases were selected for hypoglycemia warning. The sensitivity, false-positive rate and false-negative rate were used to evaluate the hypoglycemia warning performance of the LSTM-GRU model. This paper explored the changing relationship of the hypoglycemia warning performance of the model over time. Results: The predicted blood glucose levels of the three models were strongly correlated with the blood glucose levels acquired by CGMS (p < 0.001). The correlation coefficient (R-value) of the LSTM-GRU model remained stable over time (R = 0.995), nevertheless, a reduction in the R-value of the LSTM and GRU models when the Prediction Horizon (PH) was 30 min or longer. When PH was 15min, 30min, 45min and 60min, the mean RMSE values of the LSTM-GRU model were 0.259, 0.272, 0.275 and 0.278 (mmol/l), respectively, which were lower than the LSTM and GRU models and the RMSE values were statistically significant (p < 0.001). The EGA results showed the LSTM-GRU model had the highest proportion in zones A and B, as the PH extended. When PH was 30min or longer, the sensitivity and false-negative rate of the hypoglycemia warning of the LSTM-GRU model had subtle changes and the false-positive rate remained stable over time. Conclusions: The LSTM-GRU model demonstrated good performance in blood glucose prediction and hypoglycemia warning.