Brno University of Technology Smartphone PPG Database (BUT PPG)

Abstract

Brno University of Technology Smartphone PPG Database (BUT PPG) is a database created by the cardiology team at the Department of Biomedical Engineering, Brno University of Technology, for the purpose of evaluating PPG quality and estimation of heart rate (HR). The data comprises 3,888 10-second recordings of PPGs and associated ECG signals used for determination of reference HR. All of the signals contain QRS complex positions, most signals contain moreover accelerometric data (ACC), and single-value annotations of blood pressure, blood oxygen saturation and glycaemia. The data were collected from 50 subjects (25 female, 25 male) aged between 19 to 76 years at rest and during various types of movement. Recordings were carried out between August 2020 and December 2021. PPG data were collected by smartphones Xiaomi Mi9 and Huawei P20 Pro with sampling frequency of 30 Hz. Reference ECG signals and ACC data were recorded using a mobile recorder Bittium Faros 360 or 180 with a sampling frequency of 1,000 Hz (ECG) and 100 Hz (ACC). Each PPG signal includes annotation of quality and reference HR. PPG signal quality is indicated binary: 1 indicates good quality for HR estimation, 0 indicates signals where HR cannot be detected reliably and thus these signals are unsuitable for any analysis.

Background

This database was primarily created for the purpose of developing and evaluating algorithms designed to assess the quality of PPG records and algorithms designed for estimation of HR from PPG. Evaluation of PPG quality and estimation of HR from PPG have become a popular research topics [1-5], driven in part by the increased use of smartphones for health monitoring. A majority of HR estimation smartphone applications are neither tested, nor certified and moreover are not verified by medical experts. There exist some exceptions which are primarily intended for atrial fibrillation detection, e.g. Preventicus Heartbeats, FibriCheck, or Happitech. It is logical for monitoring algorithms or applications to carry out signal quality estimation before HR and other features estimation, and this approach may lead to greater robustness and reliability. For example, if a poor PPG signal quality is detected by a monitoring device, the signal may be discarded (preventing distorted and unreliable results) and a new measurement may be requested.

Methods

We recorded 3,888 10-second PPG signals using the Xiaomi Mi9 and Huawei P20 Pro smartphones. The measured subjects put their 1) index finger to the rear side of the smartphone to cover the camera and the lit LED light or 2) ear to the front camera of the smartphone in the telephoning position. The video of the finger/ear was recorded. About half of the PPG signals are from ear and half from finger. Concurrently, the single-lead ECG signal was recorded as a reference for HR calculation. In most cases, 3-axes ACC was recorded simultaneously. The Bittium Faros 360 or 180 device was used for this purpose. The electrodes were attached to the chest according to the device manual after the skin was properly prepared. PPG and ECG signals were synchronized manually. The signals were recorded even at rest or during various activities (higher pressure of finger/ear on camera, moving the finger/ear on the lens, walking, coughing, laughing, changing the light, talking). SpO2 was measured using pulse oximeter Gima, BP and glycaemia were measured by Fora D40 and Fora Diamond Mini, respectively.

Initial part of each PPG signal was removed to eliminate signal fluctuations. From the video, only the red color channel [8] was considered for annotations and the PPG signal was created using averaging of each frame. Finally, the PPG signal was inverted. The signals were divided into 10-second segments. The sampling frequency was 30 Hz (frames per second). The sampling frequency of ECG signal was 1,000 Hz and the sampling frequency of ACC signal was 100 Hz. In the ECG signal, the QRS complexes were found using robust QRS detector based on combination of three independent methods [6] and manually verified. Then, the median HR was calculated (one number for the whole 10-second signal) and this number was considered the reference HR for PPG signal.

The database is balanced in terms of gender (25 female records and 25 male records) and age (19 to 76 years, 33±17 years). The database contains binary signal-quality labels:

• 1: indicates good quality for HR estimation
• 0: indicates poor quality; signals where the HR cannot be detected reliably and thus unsuitable for analysis.

Quality annotations are based on HR estimation only from PPG signal and comparison with reference ECG value of HR. Each PPG signal was annotated by 3 or 5 experts in terms of HR. Annotators were provided with tailor-made software for the purpose of annotating, but use of the software was voluntary. The software included the picture of original PPG signal, decomposition of the PPG signal using stationary wavelet transform, HR estimated from each frequency band, spectrum created using fast Fourier transform, HR estimated from spectrum, and ruler. Annotators had the opportunity to decline to estimate the HR when the estimate would not be reliable. Annotators saw only the PPG signal (ECG signals were not provided).

Once the annotations were complete, we attempted to reach consensus for the 3 or 5 annotators. Only the "good" annotations were considered, meaning those HR values that differed from the reference HR by less than or equal to 5 bpm. The maximal error value of 5 bpm is based on the international standard IEC 60601-2-27 [7], moreover in this database it is more strict. When most annotators (2 out of 3 or 3 out of 5) provided good annotations, the signal was considered of 1 ("good quality"). When this condition was not met, the signal was labeled 0 ("poor quality").

This database was measured in two separate sessions in different time periods. All human studies were approved by the Institutional Review Board of DBME Faculty of Electrical Engineering and Communication, Brno University of Technology, on July 27, 2018 (IRB Protocol EC:EK:05b/2018) and on September 30, 2021 (IRB Protocol EC:EK:06b/2021). Informed written consent was obtained from all subjects prior to the studies.

Data Description

Each record contains PPG signal and one-lead ECG recorded with a sampling frequency of 30 Hz and 1,000 Hz, respectively. Signals from 112001 on include ACC data. All data are provided in the WaveForm Database (WFDB) format. The names (IDs) of the recordings are six-digit numbers where the first three numbers are unique subject identifiers and the next three numbers indicate the measurement number of this subject. The PPG, ECG and ACC signals are in separate files: *_PPG.dat, *_PPG.hea, *_ECG.dat, *_ECG.hea, and *_ACC.dat, *_ACC.hea, respectively.

The annotations (quality-hr-ann.csv) are recorded in a CSV file with three columns. The first column contains signal IDs. The second column contains a binary quality indicator: 1 ("good quality"), 0 ("poor quality"). The third column contains the reference HR.

QRS complex positions are within the file *_.qrs in each folder.

File subject-info.csv provides the following information:

• patient demographics (gender, age, height, weight)
• measurement spot: 0 (ear) or 1 (finger)
• motion information: 0 (rest), 1 (higher pressure of finger/ear on camera), 2 (moving the finger/ear on the lens), 3 (walking), 4 (coughing), 5 (laughing), 6 (changing the light), 7 (talking)
• physiological references: blood pressure, glycaemia, blood oxygen saturation (SpO2)

Usage Notes

The BUT PPG database is primarily intended for the development, evaluation and objective comparison of algorithms designed to assess the quality of (smartphone) PPG records (two classes) and algorithms designed for PPG HR estimation. One of the unique features of the database is that the quality annotations of each signal is based on 3 or 5 experts' annotations and their consensus, meeting international standard guidelines, and moreover meeting the stricter variation.

Secondary, the database can be used for developing and/or testing QRS complex detection algorithms. Thanks to the annotations of BP, glycaemia and SpO2, it can be used for estimation of these parameters from PPG. Moreover, the ACC signals can be used for ECG/PPG quality estimation.

Release Notes

This is the second release of BUT PPG DB. The first one contained signals 100001 to 111004. In the second release, the signals 112001 onwards were added. This new part contains additionally ACC data and single-value annotations of blood pressure, glycaemia, and SpO2. The positions of QRS complexes were added to all signals (the first release and the extension).

Ethics

This database was measured in two separate sessions in different time periods. All human studies were approved by the Institutional Review Board of DBME Faculty of Electrical Engineering and Communication, Brno University of Technology, on July 27, 2018 (IRB Protocol EC:EK:05b/2018) and on September 30, 2021 (IRB Protocol EC:EK:06b/2021). Informed written consent was obtained from all subjects prior to the studies.

Acknowledgements

This work has been funded by the United States Office of Naval Research (ONR) Global, award numbers N62909-19-1-2006 and N62909-23-1-2050.

Conflicts of Interest

The authors have no conflict of interest.

References

1. Siddiqui, S. A., Zhang, Y., Feng, Z., & Kos, A. (2016). A Pulse Rate Estimation Algorithm Using PPG and Smartphone Camera. Journal of Medical Systems, 40(5). doi:10.1007/s10916-016-0485-6.
2. Orphanidou, C. (2018). Signal Quality Assessment in Physiological Monitoring State of the Art and Practical Considerations. Cham: Springer. doi:10.1007/978-3-319-68415-4.
3. Naeini, E. K., Azimi, I., Rahmani, A. M., Liljeberg, P., & Dutt, N. (2019). A Real-time PPG Quality Assessment Approach for Healthcare Internet-of-Things. Procedia Computer Science, 151, 551-558. doi:10.1016/j.procs.2019.04.074.
4. Nemcova, A., Jordanova, I., Varecka, M., Smisek, R., Marsanova, L., Smital, L., & Vitek, M. (2020). Monitoring of heart rate, blood oxygen saturation, and blood pressure using a smartphone. Biomedical Signal Processing and Control, 59. doi:10.1016/j.bspc.2020.101928.
5. Tabei, F., Zaman, R., Foysal, K. H., Kumar, R., Kim, Y., & Chong, J. W. (2019). A novel diversity method for smartphone camera-based heart rhythm signals in the presence of motion and noise artifacts. Plos One, 14(6). doi:10.1371/journal.pone.0218248.
6. Smital, L., Marsanova, L., Smisek, R., Nemcova, A., & Vitek, M. (2020, September). Robust QRS Detection Using Combination of Three Independent Methods. In Computing in cardiology 2020.
7. International Electrotechnical Commission. (2014). Medical electrical equipment. Particular requirements for the basic safety and essential performance of electrocardiographic monitoring equipment (IEC 60601-2-27).
8. Peng, R., Zhou, X., Lin, W., & Zhang, Y. (2015). Extraction of Heart Rate Variability from Smartphone Photoplethysmograms. Computational and Mathematical Methods in Medicine, 2015, 1-11. doi:10.1155/2015/516826.