Photoplethysmography (PPG) Dataset for Non-Invasive Blood Glucose Monitoring

Abstract This project introduces an open-access database of photoplethysmography (PPG) signals annotated with reference capillary blood glucose measurements.. While PPG is widely used for heart rate monitoring, its application in metabolic sensing is a growing area of research that suffers from a lack of quality, publicly available data. Our database comprises 125 signal samples from 24 volunteers. To ensure a wide-ranging glycemic profile, data was collected from 21 healthy controls and 3 individuals with Type 1 Diabetes Mellitus (T1DM). The resulting dataset accounts for a broad metabolic spectrum, ranging from normal glucose levels to the significant glycemic excursions associated with diabetes. Data collection was performed with the in house developed multichannel ePPG device, operating across three different wavelengths (530 nm, 655 nm, 940 nm) through a 6-channel configuration. To complement the PPG data, tri-axial accelerometry data were acquired simultaneously too. To facilitate use in deep learning and signal processing research, the data is provided in wfdb format. Background Effective management of diabetes mellitus—a condition with rising global prevalence [1]—necessitates daily monitoring of blood glucose levels. Non-invasive monitoring of glycemia has gained attention within the scientific community as an alternative to current diagnostic standards, which rely on invasive or semi-invasive measurements that diminish patient comfort [2]. Various diagnostic modalities have emerged as potential solutions for non-invasive glycemic monitoring, including photoplethysmography, electrochemical impedance spectroscopy, reverse iontophoresis, and microwave-based sensing [3, 4]. Among the most promising approaches is PPG, a suitable modality for non-invasive glucose estimation due to its ability to capture the physiological impact of blood glucose on the peripheral circulatory system. Moreover, owing to its technical simplicity and compatibility with wearable technology, PPG offers a discreet alternative for continuous glycemic monitoring [2, 5]. Extracting relevant diagnostic features from the PPG waveform requires sophisticated algorithmic frameworks capable of processing optical signals and establishing a reliable correlation with blood glucose levels [6]. In recent years, artificial intelligence and machine learning have found increasingly broad applications in this domain, providing the tools necessary to decode non-linear relationships between PPG and glycemia values [7]. Developing a robust artificial intelligence model for mapping PPG signals to blood glucose levels necessitates the availability of comprehensive, open-access datasets. Such data are essential both for model training and for establishing standardized benchmarks to compare diverse architectural approaches [8]. The objective of this PPG database, annotated with corresponding blood glucose levels, is to provide the scientific community with open-access data. By doing so, it facilitates the development and validation of artificial intelligence models specifically designed for non-invasive glycemic monitoring. Methods Data acquisition was performed using the ePPG, a custom-engineered wearable platform. The device was designed and fabricated at the Faculty of Electrical Engineering and Information Technology, Slovak University of Technology (FEI STU), to facilitate multi-sensor physiological monitoring [9]. The ePPG device recorded all six PPG channels simultaneously at a sampling rate of 100 Hz. To facilitate motion artifact characterization, PPG data capture was synchronized with a built-in three-axis accelerometer, which monitored physical movement across all three spatial axes throughout the recording process. Reference blood glucose levels were established via capillary fingerstick measurements using a blood glucose monitor GL 49 Beurer glucometer on the same hand as the ePPG device. Following hand sanitization and fingertip disinfection, a blood sample was collected for the initial readout. The ePPG device was then secured to the left thumb. To ensure data integrity, signal recording commenced only after the PPG waveform reached a stable, artifact-free state, with this stabilization phase strictly limited to a maximum of one minute. The duration for each following PPG signal recording was at least 30 seconds. Data Description The dataset comprises 125 records obtained from 24 distinct subjects. Each record consists of six PPG channels captured using two photodetectors across three wavelengths [9]: green (530 nm), red (655 nm), and infrared (940 nm). These signals are labeled G1, R1, IR1 and G2, R2, IR2, where the letter indicates the wavelength (G, R, or IR), and the numerical suffix identifies the specific detector. Signal values are expressed in arbitrary units (a.u.). Additionally, each record includes synchronized triaxial accelerometer data (X, Y, and Z components) for motion tracking. The dataset is structured in the WFDB format, where each record is comprised of three distinct files: a header file with the .hea extension, a binary data file with the .dat extension, and an annotation file with the .atr extension. Individual records are organized into multiple directories labelled Subject_xyz in root directory. Each directory contains all records acquired from a single volunteer. The number of records per participant is not uniform, and the duration of individual recordings is trimmed to 30 seconds. In general, the records was administered at various times throughout the day—encompassing pre-prandial, post-prandial, and inter-meal intervals—without dietary restrictions. In cases where multiple measurements were obtained per subject, these sessions were conducted across several different days. Demographic and clinical data for each participant are stored in a file named subjects.csv, located in the root directory. For each subject, the recorded attributes include a unique identifier (ID), age, sex (F = female, M = male), height, weight, and the presence of T1DM. Within the subject cohort, 3 individuals were diagnosed with Type 1 diabetes, while the remaining participants were normoglycemic. This demographic and clinical data is also additionally incorporated into the header file (.hea) of each respective record. The annotation file (.atr) contains blood glucose levels (measured in mmol/L), which were manually recorded during the session. These reference measurements were obtained immediately prior to the PPG acquisition process (maximum with 1min delay), ensuring that glycemia data was collected shortly after each PPG recording for temporal alignment. In the annotation file (.atr), each annontation is timestamped to coincide with the first sample of the recording. Quality-Control Annotations In addition to the raw recordings, the record distributes two CSV files with the results of the signal quality control described in the accompanying data article: qc_records.csv - one row per record, with record-level quality verdicts: the number of PPG channels passing quality control, the best channel’s fraction of analysable 8-second windows, median heart rate across passing channels, an accelerometer-derived motion index, and the longest run of consecutive clean pulses together with its channel and start/end times in seconds, allowing clean segments to be extracted directly without re-running any analysis. qc_channels.csv - one row per record–channel pair (750 rows), with per-channel signal quality metrics: perfusion index, skewness and relative-power signal quality indices, detected heart rate and inter-beat-interval variability, fraction of analysable windows, beat-template correlation with template-validity flag, and the longest clean pulse run per channel. Quality control was computed on the 0.5–8 Hz band-passed signals in 8-second windows advanced by 2 seconds, with beat-level assessment by correlation against each record's own median beat template (threshold 0.80), applied only where the template is itself pulse-shaped. Usage Notes There are few publicly available PPG datasets accompanied by well-curated blood glucose annotations. Our primary objective in developing this database was to address this scarcity. We anticipate that this dataset will be instrumental for the development and objective benchmarking of non-invasive glycemia estimation algorithms. Records failing quality control are deliberately retained to reflect realistic wearable acquisition conditions and to support research on signal quality itself; they can be excluded with a single filter on the record_passes or best_clean_run columns of qc_records.csv. Ethics The study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was reviewed and approved by the Ethics Committee of the F. D. Roosevelt Faculty Hospital with Polyclinic in Banská Bystrica, Slovakia (approval No. 18/2021). All participants provided written informed consent prior to enrollment. All data were de-identified in accordance with institutional and ethical guidelines before public release. Acknowledgments This work was funded by the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project No. 09I05-03-V02-00061.