Abstract
Continuous monitoring of glucose levels is important for diabetes management and prevention. While traditional glucose monitoring methods are often invasive and expensive, recent approaches using machine learning (ML) models have explored non-invasive alternatives-but many still depend on manually logged food intake and activity, which is burdensome and impractical for everyday use. In this study, we propose a novel approach that eliminates the need for manual input by utilizing only passively collected, automatically recorded multi-modal data from non-invasive wearable sensors. This enables practical and continuous glucose prediction in real-world, free-living environments. We used the BIG IDEAs Lab Glycemic Variability and Wearable Device Data (BIGIDEAs) dataset, which includes approximately 26,000 CGM readings, simultaneous ly collected wearable data, and demographic information. A total of 236 features encompassing physiological, behavioral, circadian, and demographic factors were constructed. Feature selection was conducted using random-forest-based importance analysis to select the most relevant features for model training. We evaluated the effectiveness of various ML regression techniques, including linear regression, ridge regression, random forest regression, and XGBoost regression, in terms of prediction and clinical accuracy. Biological sex, circadian rhythm, behavioral features, and tonic features of electrodermal activity (EDA) emerged as key predictors of glucose levels. Tree-based models outperformed linear models in both prediction and clinical accuracy. The XGBoost (XR) model performed best, achieving an R-squared of 0.73, an RMSE of 11.9 mg/dL, an NRMSE of 0.52 mg/dL, a MARD of 7.1%, and 99.4% of predictions falling within Zones A and B of the Clarke Error Grid. This study demonstrates the potential of combining feature engineering and tree-based ML regression techniques for continuous glucose monitoring using wearable sensors.