Assessment of Deep Learning Model System for Blood Glucose Time-Series Prediction

Abstract

Diabetes has become one of the most severe and prevalent chronic diseases, leading to life-threatening, costly, and disabling consequences and reduced life expectancy. Uncontrolled blood glucose (BG)  conditions become a factor in diabetes mellitus sufferers, which then causes BG levels that are too high (hyperglycemia) and too low (hypoglycemia). People with Type 1 Diabetes Mellitus (T1DM) require long-term BG management to keep BG levels. Deep learning models using Continuous Glucose Monitoring (CGM) data to monitor and regulate BG concentrations in diabetic patients with prediction values to prevent hypoglycemia and hyperglycemia is very important. Based on some of the latest research, the deep learning Temporal Fusion Transformer (TFT) model is considered an approach method with superior performance in time-series prediction. Therefore, in this study, two TFT models, the TFT and AutoTFT univariate models, were proposed for the time-series BG prediction for T1DM patients. In this study, the two proposed TFT models with two baseline models, were trained and tested on the ShanghaiT1DM dataset. The proposed and baseline models were trained using manual and auto-tuning hyperparameters with Optuna on cross-validation for prediction horizons (PHs) of 30 and 60 minutes, respectively. The performance metrics used to evaluate the models were mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE). As a result, the TFT model is superior to the baseline LSTM model, also the proposed AutoTFT models achieved the smallest MAE, MAPE, and RMSE  for both 30 and 60-minute PHs, respectively of all models used. Besides, the BG prediction results with 30-minute PHs are better than those with 60-minute PHs for all the models. This shows that the AutoTFT model stands as a promising tool for the accurate prediction of adverse glycemic events.