Impact of Teacher Forcing Training on Blood Glucose Prediction Performance

Abstract

In this study, the accuracy of blood glucose prediction was improved using the teacher forcing LSTM (TF-LSTM) method, and the model was compared with a traditional LSTM approach. The model was trained using the OhioT1DM dataset and the DMMS.R simulation program, which contain data obtained from patients with Type 1 diabetes. Models were trained with past window lengths of 15, 30, and 45 minutes, and their performance was analyzed for the corresponding prediction horizons. The findings of this study indicate that the teacher forcing LSTM method significantly reduced prediction error, particularly in short-term predictions using the OhioT1DM dataset. However, in the DMMS.R dataset, a partial decline in model performance was observed as the prediction horizon increased. The impact of individual differences on prediction performance was also analyzed, revealing that while error decreased significantly for some patients, improvements were more limited for others. Overall, these findings demonstrate that the TF-LSTM method is effective in blood glucose level prediction, especially for short-term prediction horizons.