PhysioCGM: a multimodal physiological dataset for non-invasive blood glucose estimation

All data records in the dataset are included in this submission and are accessible via FigShare. Table 3 in the manuscript provides a detailed overview of these records. The dataset is organized into two cohort folders, each containing data from five participants. Each subject is identified by a de-identified code, e.g. c1s01 represents the first subject from cohort 1. Within each subject’s folder, there are three subfolders: zephyr, e4, and CGM, each storing the data from respective sensors. The zephyr folder contains session folders of recordings, with each session folder including seven CSV files that capture various data types, including ECG, acceleration, and breathing waveforms. The e4 folder stores data from the Empatica E4 watch, including PPG, EDA and accelerometry. Lastly, the CGM folder contains a file that stores glucose values. Each data record includes a timestamp field, which allows time alignment of signals from different sensors. To streamline data access and training, we preprocess the raw data records from all sensors and align them 102 based on their corresponding timestamps. They are then split at the CGM level and packed into pkl binary 103 files for ease of use. The processed data is organized by subjects. Within each subject’s folder, there are 104 multiple subfolders corresponding to different days, each of which contain multiple pkl files. Each pkl file 105 contains 5-minute signal clips and metadata that are synchronized with CGM timestamps.