SugarNet: Personalized Blood Glucose Forecast for Diabetic Patients with Joint Time-Frequency Learning

Abstract

Diabetes is a chronic condition characterized by ineffective regulation of blood glucose (BG). Foreseeing upcoming BG trends is crucial to tight glycemic control; however, existing models for BG prediction only use time domain signals. This paper introduces SugarNet, a novel deep learning model for personalized BG forecasts based on food intake, medication, insulin, and BG history. The model augments the inputs with an embedding block aiming for a more expressive representation of the signals. Both the augmented time series and its frequency spectrum are forwarded to two dedicated multi-layer convolution blocks, followed by Long Short-Term Memory (LSTM) networks. The outputs are then fused to generate the forecasts of differences in BG values between current time T and future time T'. Along with novel feature engineering techniques, the model is pre-trained on a set of patients and then fine-tuned and tested on a different set of patients with transferred knowledge. Extensive experiments on 12 type 1 diabetes patients and 100 type 2 diabetes patients yielded improved RMSE by 14.7%-21.8%, demonstrating the superiority of the model over state-of-the-art methods. By providing accurate insights into future BG values, SugarNet can significantly enhance diabetics’ management of their condition.