Developing Consumer Wearable Compatible Computational Models for Non-Invasive Glucose Estimation in Diabetes Patients

Regular blood glucose monitoring is crucial for managing Type 1 diabetes (T1D). Traditional finger-prick tests and current continuous glucose monitoring (CGM) systems have notable limitations, including being invasive and costly. This study aims to develop computation models that continuously estimate glucose levels using multimodal data readily collected from consumer wearable devices in everyday life. We used the T1D subset of the D1NAMO dataset, which includes simultaneous glucose readings and wearable data. Rigorous feature extraction was conducted to identify four main categories of features likely related to glucose levels, as indicated by previous research. The most relevant features were then selected to train and tune seven machine learning (ML) regression models. Tree-Based methods demonstrated good performance in terms of data and clinical accuracy. The top-performing model, Light Gradient Boosting Regression, achieved an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.795, NRMSE of 0.452 mg/dL, MARD of 15.3%, and placed 97.7% of predictions within clinically acceptable zones of the Clarke Error Grid. Our models hold potential for large-scale adoption due to their compatibility with widely available consumer wearables.