Selection of Noninvasive Features in Wrist-Based Wearable Sensors to Predict Blood Glucose Concentrations Using Machine Learning Algorithms

Abstract

Glucose monitoring technologies allow users to monitor glycemic fluctuations (e.g., blood glucose levels). This is particularly important for individuals who have diabetes mellitus (DM). Traditional self-monitoring blood glucose (SMBG) devices require the user to prick their finger and extract a blood drop to measure the blood glucose based on chemical reactions with the blood. Unlike traditional glucometer devices, noninvasive continuous glucose monitoring (NICGM) devices aim to solve these issues by consistently monitoring users' blood glucose levels (BGLs) without invasively acquiring a sample. In this work, we investigated the feasibility of a novel approach to NICGM using multiple off-the-shelf wearable sensors and learning-based models (i.e., machine learning) to predict blood glucose. Two datasets were used for this study: (1) the OhioT1DM dataset, provided by the Ohio University; and (2) the UofM dataset, created by our research team. The UofM dataset consists of fourteen features provided by six sensors for studying possible relationships between glucose and noninvasive biometric measurements. Both datasets are passed through a machine learning (ML) pipeline that tests linear and nonlinear models to predict BGLs from the set of noninvasive features. The results of this pilot study show that the combination of fourteen noninvasive biometric measurements with ML algorithms could lead to accurate BGL predictions within the clinical range; however, a larger dataset is required to make conclusions about the feasibility of this approach.

Keywords: blood glucose self-monitoring; continuous monitoring; diabetes mellitus; glucose; hyperglycemia; machine learning; noninvasive.

Figure 1

Sensors and their features for the (a) OhioT1DM and (b) UofM datasets.

Figure 2

Flowchart of the machine learning pipeline for evaluating different regression models to estimate blood glucose values from a set of noninvasive features.

Figure 3

The correlation coefficient between target glucose concentrations and selected features of subject (a) 559 and (b) 563 from the OhioT1DM dataset.

Figure 4

Performance of the best-trained model for both subjects (a,b for Subject 559 and c,d for Subject 563) from the OhioT1DM dataset. In panels (a,c), the predicted and target glucose concentrations are sorted in ascending order. The Clarke Error Grids (reference versus predicted glucose values) are shown in panels (b,d). The Clarke Error Grid separates the measurements into five regions based on their accuracy; read Section 2.4 for more details.

Figure 5

The correlation coefficient between target glucose concentrations and the selected fourteen features for both subjects, (a) Subject 1 and (b) Subject 2, from the UofM dataset.

Figure 6

Performance of the best-trained model in unseen instances for both subjects from the UofM dataset. Panels (a,c) show the predicted and target glucose concentrations from unseen data. The Clarke Error Grids are shown in panels (b,d). RMSE and R² values of the best model are reported in panels (b,d). We also report the percentage of instances (red font) falling in each region of the Clarke Error Grid.