A multimodal physiological dataset for non-invasive blood glucose estimation

Diabetes is a major health challenge that affects millions of people worldwide. Managing diabetes effectively requires monitoring blood glucose levels continuously, typically through invasive sensing devices such as continuous glucose monitors (CGMs). Blood glucose excursions have been shown to induce changes in several physiological signals such as electrocardiography (ECG), photoplethysmography (PPG) and electrodermal activity (EDA) that can be measured non-invasively with consumer-grade wearable sensors. These physiological changes can be mapped into glucose levels using machine-learning models trained on comprehensive multimodal datasets of physiological signals. However, most existing datasets lack ground-truth measurements from CGMs and often only include aggregated physiological data (e.g. heart rate) at low sampling rates. To address these gaps, we introduce PhysioCGM, an open-source dataset that contains raw physiological recordings from multiple sensors including ECG, PPG, EDA, skin temperature, accelerometry and ground-truth CGM data collected for up to 17 days from 10 participants with Type 1 Diabetes in ambulatory settings. This dataset aims to promote the development of non-invasive methods for glucose monitoring and improve diabetes management.