Deep Learning-Based Glucose Prediction Models: A Guide for Practitioners and a Curated Dataset for Improved Diabetes Management

Abstract

Accurate short- and mid-term blood glucose predictions are crucial for patients with diabetes struggling to maintain healthy glucose levels, as well as for individuals at risk of developing the disease. Consequently, numerous efforts from the scientific community have focused on developing predictive models for glucose levels. This study harnesses physiological data collected from wearable sensors to construct a series of data-driven models based on deep learning approaches. We systematically compare these models to offer insights for practitioners and researchers venturing into glucose prediction using deep learning techniques. Key questions addressed in this work encompass the comparison of various deep learning architectures for this task, determining the optimal set of input variables for accurate glucose prediction, comparing population-wide, fine-tuned, and personalized models, and assessing the impact of an individual's data volume on model performance. Additionally, as part of our outcomes, we introduce a meticulously curated dataset inclusive of data from both healthy individuals and those with diabetes, recorded in free-living conditions. This dataset aims to foster research in this domain and facilitate equitable comparisons among researchers.

Keywords: Diabetes; Glucose prediction; deep learning; transfer learning.

Figure 1.

Monitoring devices used in the study.

Figure 2.

Processed signals of a representative subject with T1DM: (a) CGM; (b) IGAR and Plasma Insulin, and (c) Heart Rate and Number of Steps.

Figure 3.

Inference process common to all models. The architecture depicted includes all the available variables, which can vary based on the selected group (see Table 2). Within each n-dimensional sliding window of length , input variables are normalized across the time dimension. A preprocessing step involving the feature extraction layer is applied to the inputs, which are then processed by the primary network to produce a one-step-ahead prediction. This prediction is added to the moving window as if it were a new measurement. Exogenous variables are propagated as described in Section II-C. This iterative process is repeated until the requested predictions are generated. The predicted sequence is then denormalized using parameters and , obtained during the normalization.

Figure 4.

Results for the healthy and T1DM population. Here pop, pers, and fine, stand for the population, personalized, and fine-tuning training approaches, respectively.

Figure 5.

Clarke Error Grids representing the clinical significance of predicted versus actual blood glucose (BG) values, categorized into five distinct zones A to E, aiding in the visual assessment of glucose prediction accuracy and potential clinical implications. (a) patients with T1DM, (b) healthy subjects.