Predicting the Blood Glucose Level Using Transformers

In this paper, we explore the use of Transformer, a novel deep learning model, to predict blood glucose levels in type 1 diabetes patients. We also evaluate the impact of two preprocessing steps: data imputation and smoothing. We benchmark the Transformer against two other models-XGBoost and a one-dimensional convolutional neural network (1D-CNN)-using the OhioT1DM dataset to forecast blood glucose levels 30 minutes ahead based on a 60-minute historical timeframe. Additionally, we assess four methods for addressing missing time series data during training: hourly mean, linear interpolation, cubic interpolation, and spline interpolation, along with two smoothing approaches: Kalman smoothing and smoothing splines. Our tests indicate that the Transformer outperforms both XGBoost and 1D-CNN when relying solely on continuous glucose monitoring (CGM) as a predictor. Moreover, it holds its own against XGBoost when both CGM and meal-induced carbohydrate intake are predictors. Our findings demonstrate higher accuracy than previously reported in existing literature.