Toward Short-Term Glucose Prediction Solely Based on CGM Time Series

TL;DR

This paper introduces TimeGlu, an end-to-end pipeline for short-term glucose prediction using solely CGM time series data, addressing the need for real-time decision support without multi-modal privacy concerns. The authors compare TimeGlu against four baselines (Exponential Smoothing, Auto ARIMA, BATS, TBATS) and demonstrate superior accuracy and adaptability on the CGM Glucose and Colás datasets, using a Bi-LSTM encoder with Additive Attention and a lightweight decoder. Data augmentation via Gaussian noise and a robust encoder-decoder architecture drive strong quantitative gains (MAE and MAPE) and qualitative robustness across normal and outlier glucose values. The work supports real-world diabetes management by providing accurate, privacy-preserving short-term glucose forecasts that generalize across diverse patient populations and recording conditions.

Abstract

The global diabetes epidemic highlights the importance of maintaining good glycemic control. Glucose prediction is a fundamental aspect of diabetes management, facilitating real-time decision-making. Recent research has introduced models focusing on long-term glucose trend prediction, which are unsuitable for real-time decision-making and result in delayed responses. Conversely, models designed to respond to immediate glucose level changes cannot analyze glucose variability comprehensively. Moreover, contemporary research generally integrates various physiological parameters (e.g. insulin doses, food intake, etc.), which inevitably raises data privacy concerns. To bridge such a research gap, we propose TimeGlu -- an end-to-end pipeline for short-term glucose prediction solely based on CGM time series data. We implement four baseline methods to conduct a comprehensive comparative analysis of the model's performance. Through extensive experiments on two contrasting datasets (CGM Glucose and Colas dataset), TimeGlu achieves state-of-the-art performance without the need for additional personal data from patients, providing effective guidance for real-world diabetic glucose management.

Figures (6)

• Figure 1: Overview of the proposed TimeGlu pipeline. Noise with standard normal distribution is applied to the input data for data augmentation, which is then input into the encoder-decoder-based architecture. To precisely and comprehensively extract the time series features, a sequential Bi-LSTM structure with an Additive Attention bahdanau2014neural module is designed. The high-dimensional features will be input into a lightweight decoder to generate the predicted glucose value. An MSE loss is constructed for pipeline training.
• Figure 2: Visualization of the CGM Glucose dataset.Left: Three different ranges are demonstrated across days within one week. Right: TIR/TAR/TBR visualization of the dataset. Multiple levels (median, min, max, etc.) of glucose values are illustrated.
• Figure 3: Visualization of the Colás dataset.Left: 30 subjects with the most abnormal blood glucose situations are shown as an example. Right: 30 randomly selected subjects are illustrated through the box plot to showcase the variety and data diversity of the dataset.
• Figure 4: Visualization of glucose value prediction of TimeGlu on the CGM Glucose dataset. TimeGlu is capable of accurately predicting glucose trends and specific values over time on the CGM Glucose dataset.
• Figure 5: Visualization of glucose value prediction of TimeGlu on the Colás dataset. For more clear demonstration, 114 data points are illustrated in the figure. For large-scale datasets, TimeGlu provides robust generalization capabilities to accurately predict blood glucose trends as well as glucose values.