A Leakage-Controlled Evaluation of Multimodal Sensor Fusion for Wrist-Worn Glucose Estimation

A study evaluated the accuracy of wrist-worn wearable devices for glucose measurement using multimodal sensors. The research involved 15 participants with continuous glucose monitoring (Dexcom G6) and an Empatica E4 wristband. When forecasting glucose levels 30 minutes ahead from continuous monitoring history, all tested models (linear regression, gradient-boosted trees, and neural networks) achieved similar accuracy with RMSE error of 13.90 mg/dL, indicating an information ceiling. Adding data from photoplethysmography, electrodermal activity, skin temperature, and accelerometry did not improve accuracy. Most significantly, glucose estimation using only wristband data (22.58 mg/dL) was statistically indistinguishable from a model using only time of day (22.63 mg/dL). The research concluded that in this study group, wristband signals contain no glucose information beyond the group mean.