Database Open Access
Continuous Cuffless Monitoring of Arterial Blood Pressure via Graphene Bioimpedance Tattoos
Published: June 4, 2022. Version: 1.0.0
When using this resource, please cite:
(show more options)
Ibrahim, B., Kireev, D., Sel, K., Kumar, N., Akbari, A., Jafari, R., & akinwande, d. (2022). Continuous Cuffless Monitoring of Arterial Blood Pressure via Graphene Bioimpedance Tattoos (version 1.0.0). PhysioNet. https://doi.org/10.13026/ce62-pc98.
Please include the standard citation for PhysioNet:
(show more options)
Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P. C., Mark, R., ... & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220.
Continuous monitoring of arterial blood pressure (BP) in non-clinical (ambulatory) settings is essential for a proper understanding of numerous health conditions. While conventional ambulatory BP devices exist, they are uncomfortable, bulky, and intrusive. In this work, we introduce a dataset of BP measurement using atomically-thin graphene electronic tattoos (GETs) and bioimpedance (Bio-Z) as measurement modality. The Bio-Z is measured from radial and ulnar arteries on the wrist. Afterward, a machine learning regression model is employed to relate the recorded bioimpedance to BP, yielding an effective beat-to-beat BP monitoring platform. In this database, we are sharing the raw time data for the 4-channel GET Bioimpedance signals with BP and PPG signals. The GETs are used to monitor arterial blood pressure from N=7 individuals for 300+ minutes each. A medical-grade BP monitoring device, Finapres NOVA, was used to measure the subject's control BP during the experiments. On average, each participant underwent 4±1 hours of continuous BP monitoring. Altogether, we performed on average ~2500±600 BP measurements per participant, which resulted in a total of 18667 datapoints.
Monitoring blood flow is perhaps among the most basic practices of modern medicine. Yet, it is known that cardiovascular diseases such as sleep apnea, stroke, or hypertension are still on rise and it is essential to monitor patient’s BP continuously and routinely –. Medical doctors rely on cuff sphygmomanometers, measuring single values of systolic (SBP), diastolic (DBP), and mean arterial (MAP) blood pressures –.
Directly capturing an individual’s blood pressure in a continuous manner is a non-trivial technological challenge. A few cuff-less BP monitoring methods exist, relying on optical , , acoustic , , or pressure ,  modalities. The common drawbacks of all these systems are their bulkiness and incompatibility with skin’s elastic properties. Unlike other works to date, the graphene enabled BP and Bio-Z technology presented here is equally operational regardless of the skin color and able to perform long-term and nocturnal measurements.
Graphene Tattoo Fabrication
Graphene tattoo are fabricated on ultrathin, 200 nm layer of polymethyl methacrylate (PMMA) and using high-quality CVD graphene .
Graphene-Skin Impedance Measurements
The graphene-skin impedance is studied via the Hioki LCR meter IM3536.
Bio-Z Measurement Setup
The low-noise multichannel Bio-Z sensing hardware, the so-called XL-board, is explicitly designed to capture the slight variations in Bio-Z with high resolution. A custom-developed printed circuit board (PCB) is designed to provide low-noise Bio-Z sensing for this study. The hardware is built around the ARM Cortex M4 microcontroller (MCU), which transfers the user-defined digital waveform into the AC current signal by passing a 16-bit digital-to-analog converter (DAC, DAC8811, Texas Instruments, USA). The MCU controls the frequency and amplitude of the current signal. In its turn, the DAC generates an analog signal utilized into a negative feedback loop on a low-noise operational amplifier (OPA211, Texas Instruments, USA) to generate an AC current signal with programmable amplitude and frequency. A series capacitor at the DAC output is used to circumvent a DC current component's injection into the human body. The signal from the impedance sensing electrodes is filtered through the high-pass filter.
To gain the Bio-Z, we measure the voltage modulation associated with the injected current modulation. The signal is then amplified with a low-noise instrumentation amplifier (IA). A high-precision analog-to-digital converter (ADC) facilitates the IA output through an analog anti-aliasing low-pass filter. The ADC (ADS1278, Texas Instruments, USA) samples the voltage at 93.75 kHz sampling frequency with a 24-bit (0.3 μ V) resolution to provide sufficient precision. The analog front end and the MCU can handle to measure simultaneously 10 independent Bio-Z streams and various analog readings.
In this study, the first four channels are selected for high-resolution Bio-Z sensing and the fifth channel is reserved for simultaneous PPG readings used for syncing with the Finapres NOVA BP system. The sampled data is forwarded to PC via the MCU and Hi-Speed USB Bridge for signal post-processing. When recording Bio-Z signals, each subject wears additional sensors alongside the GETs: BP brachial cuff and finger cuff (Finapres NOVA) and two PPG sensors. The two PPG sensors allow us to precisely correlate the timing of the events between the XL board and the Finapres device.
Human Subject Experiments
A total of N=7 subjects in their middle twenties have participated in this proof-of-principle study. N=6 subjects (Subject 1 to 6) performed graphene tattoo experiments with HGCP and Valsalva maneuvers as BP elevation, a series of mild exercises, and post-workout HGCP model validation experiments as shown in Table 1. Subject 7 only performed graphene tattoo experiments with cycling-enabled BP elevation routine.
Each experimental routine was at least ~3 hours long in average, up to 5 hours. Each HGCP pattern itself takes ~450 seconds of data collection, and the average of 5±1 HGCP patterns were performed by each subject, comprising a total of ~2500±600 samples (beats) used for algorithm training and ~250±60 samples used for algorithm testing. Valsalva routines comprised an average of 300±150 samples, and cycling routines consisted of ~615 samples (down-sampled to ~121 samples for additional window averaging) that were considered for model training and testing.
Table 1. Summary of all the data collection experiments showing the subject number, day of data collection, duration of experiment and activities done within the experiment.
|Data Folder||Subject #||Day of Collection||Duration, min||Measurements||Post-Break Measurements||Break Activity||Comments|
|subject1_day1||1||1||180||Baseline, HGCP, Valsalva||Yes||Walk and stairs|
|subject1_day4||1||4||40||HGCP, Valsalva||No||-||After 4 days of routine activity|
|subject2_day1||2||1||170||Baseline, HGCP, Valsalva||Yes||Walk and stairs|
|subject3_day1||3||1||180||Baseline, HGCP, Valsalva||Yes||Walk and stairs|
|subject4_day1||4||1||150||Baseline, HGCP, Valsalva, Rest||Yes||Walk and stairs|
|subject5_day1||5||1||195||Baseline, HGCP, Valsalva||Yes||Extensive workout|
|subject6_day1||6||1||320||Baseline, HGCP, Valsalva||Yes||Eating a meal, sweat walk (99F)|
Bio-Z Signal Processing
The acquired signal via XL-board is band-pass filtered (second-order Butterworth) centered around the driving AC frequency to remove the residual DC offset, 60 Hz interference, and high-frequency noise. Then, Bio-Z is extracted via simultaneous demodulation by multiplying the filtered signal by the injection signal generated by the MCU. The multiplier output is low-pass filtered (second-order Butterworth) with the cut-off frequency of 6 Hz to remove the carrier signal distortion and out-of-band noise, yet allowing us to measure extreme maximum heart rates. Prior to the operation, the hardware is calibrated by measuring a known resistor's impedance to convert the measured voltage to an accurate resistance value. The measurement system can measure impedance with a root mean square (RMS) error less than 1 mΩ , which is much lower compared to the target Bio-Z variations.
Signal abstraction, Feature extraction, and BP Regression model
N=4 characteristic points abstract the recorded ΔZartery(t) signals for each heartbeat. The ΔZartery peak corresponds to the DBP, while the most pronounced minima correspond to the SBP. Additionally, the second smaller peak and minima in the middle of the cardiac cycle correspond to the back reflection of the pressure pulse. In addition to SBP and DBP, we leverage the aforementioned characteristic points to estimate mean arterial pressure (MAP). We extract the reference MAP from BP waveform by taking the area under the BP curve normalized by cycle duration. To effectively detect DBP, SBP, and MAP, we use four characteristic points from all phases of the cardiac pulse of the ΔZartery(t) signal: (1) diastolic peak (DIA), (2) maximum slope (MS), (3) systolic foot (SYS) and (4) inflection point (IP). The DIA and SYS are estimated via the intersection of the tangent to the slope with the horizontal line from the maximum and the minimum of the signal, respectively. MS is estimated as the point in the middle of the descending slope section. IP is the maximum slope point between the second peak and the notch. The points are identified from the first and the second derivative of the ΔZartery(t) signal using the zero crossing, peak, and foot points.
Pulse transit time, the time it takes for the pulse travel between two sensing sites on one artery is selected as one of our main features of BP. BP has an inverse quadratic relation with PTT. The ratio between the amplitudes of SYS and IP relative to the DIA is a measure of the reflection wave's intensity. The time interval between SYS and IP measures the arterial stiffness, while the area under the curve represents the total peripheral resistance. All the features mentioned above are useful in modeling the arteries' cardiovascular properties, and we use them to build our regression model for BP estimation. The characteristic points of the Zartery(t) signals mentioned above are used to generate as many as 50 features for each heartbeat. The features are categorized into four sets, such as PTT, timepoint, amplitude, and area. The PTT features are calculated from every possible pair of signals; other elements are computed from each signal individually. In order to smooth the data and filter out the beat-to-beat variations, we used window-based averaging (20 adjacent beats with a 50% overlap).
For each model, the hyper-parameters are selected by splitting the training data into 89%/11% fractions, using the 89% to estimate the model parameters, and the remaining 11% performance evaluation. AdaBoost models' hyper-parameters consist of the number of the decision trees and the tree depth, which are selected from 8 or 16 for the number of trees and 4 or 8 for the tree depth that achieves minimum error in the validation dataset.
The DBP, SBP, and MAP are finally estimated using an advanced regression model trained by the ΔZartery(t) features extracted and BP data measured simultaneously by a reference continuous BP monitoring device (Finapres, NOVA). Our method provides brachial BP values from the wrist pulse signals' features since the regression training model uses brachial BP data. To effectively estimate the BP, we require to measure multiple values of both DBP and SBP during the training cycle. Therefore, such a training cycle usually consists of a series of exercises followed by a short relaxation state. Our subject-specific models (separate for SBP and DBP) are trained using a minimal number of training window samples for each subject, bringing an additional requirement of careful model selection to avoid overfitting.
We use the Adaptive Boosting regression model, which is based on ensemble learning that builds the prediction by combining several weak learners' outputs through a weighted sum of different subsets of the training data set. The BP estimation performance is evaluated through training and testing the BP models on different subsets of the data as follows. First, The HGCP data of each subject is shuffled and then divided into ten folds to apply cross-validation by training the model using 9 folds (90% of the data) and testing the model on the remaining fold (10% of the data) for ten times by changing the testing fold each time to cover the whole data and avoid the bias for training the model with a certain part of the data. Once the training cycle and regression model are complete, we use the regression model to directly output the BP(t) from the measured ΔZartery(t).
The performance of the models is evaluated using the average across all the ten folds of the BP’s RMSE. Second, the HGCP and Valsalva data of each subject are used for model training based on 10-fold cross-validation similar to the previous case but without shuffling to test the model capability of estimating BP for a continuous time-segment. Third, each subject's Valsalva data are the only data used for model training based on 10-fold cross-validation without shuffling to measure the capability of training the model with a small amount of data. Fourth, the HGCP and Valsalva data are used only to train a single model for each subject, which is tested on the post workout data to evaluate the BP estimation in the future and after workout. Lastly, for subject 1, a model is trained by the HGCP and Valsalva data that was collected on the first day, then the model is tested on HGCP data measured after 4 days to evaluate the repeatability of BP estimation after multiple days.
Data is divided into folders based on subject number and day of experiment. Data is collected for each subject in a single day except subject 1 that has data for 2 different days (day1 and day4) to test the repeatability of blood pressure prediction after several days as shown in Table S1. Each subject data is divided into sub-folders based on setups (baseline, HGCP, ..., etc) and index to show the order of the setups. There are multiple folders for the same setup with different index which are repeated trials for the same setup. Each setup folder consists of the data files in CSV format called data_trial. Each data type is in a separate data_trial CSV file. Each data_trial CSV file for specific data type is segmented into multiple files with index 01,02,..etc to avoid large file size. The data in data_trial CSV files were collected successively after each other based on the sequence of the index. Data within the data_trial CSV file is arranged in columns. The first column is timestamps in seconds and the rest of columns are the data.
BioZ signals are in mOhms. BP signal is in mmHg. PPG signals are a.u.. Signals are time-synchronized.
- *bioz data = BioZ1 (radial artery closer to wrist), BioZ2 (radial artery closer to heart), BioZ3 (ulnar artery closer to wrist), BioZ4 (ulnar artery closer to heart). Sampling Rate = 1250 Hz
- *finapresBP = Continuous BP measurement with Finapres representing brachial BP. Sampling Rate = 200 Hz
- *finapresPPG = PPG measurement from Finapres on fingertip. Sampling Rate = 75Hz
- *ppg = PPG measurement from fingertip with BioZ XL board. Sampling Rate = 1250 Hz
- Hand Grip Cold Pressor (HGCP): the subject performs handgrip (HG) exercise for 3 minutes, slowly raising their DBP and SBP, then placing their hand into an ice cold water bucket (cold pressor, CP) for 1 minute to ensure that BP first goes even higher, then very slowly decreases over the 4 minute resting period.
- Cycling: the subject is stationed to perform a set of bike cycling treadmill exercises for 4 minutes, with 4 minutes of break for resting in between.
- Valsalva: session with multiple Valsalva maneuvers. Each Valsalva maneuver consists of a subject pinching their nose while trying to breathe out intensely for 20-30 seconds, creating an extensive buildup of inner pressure, both raising BR, then decreasing, and rapidly increasing it once again very rapidly.
- Baseline: At the beginning of data collection with no BP change protocol. Participants are at rest.
- Rest: no BP change protocol. Participants are at rest.
More information about the Finapres® NOVA device for reference blood pressure used in this experiment can be found on the company website .
The machine learning (ML) algorithms for the prediction of cuffless blood pressure based on this data from Graphene Bioimpedance Tattoos is publicly available online at GitHub .
Version 1.0: Initial release.
The human subject BP measurements were performed under the approval of the Institutional Review Board of the University of Texas A&M (IRB no. IRB2017-0335D). The tattoo characterization experiments were performed under the approval of the Institutional Review Board of the University of Texas at Austin (IRB no. 2018-06-0058).
Conflicts of Interest
The authors, RJ and BI, filed a patent related to this research and this patent is licensed to SpectroBeat LLC. The associated patent application is US 2020/0138303 with title "System and method for cuff-less blood pressure monitoring".
- H. Al Ghorani, S. Kulenthiran, L. Lauder, M. Böhm, and F. Mahfoud, “Hypertension trials update,” J. Hum. Hypertens., vol. 35, no. 5, pp. 398–409, 2021.
- O. Marrone and M. R. Bonsignore, “Blood-pressure variability in patients with obstructive sleep apnea: Current perspectives,” Nat. Sci. Sleep, vol. 10, pp. 229–242, 2018.
- M. R. Salazar et al., “Nocturnal hypertension in high-risk mid-pregnancies predict the development of preeclampsia/eclampsia,” J. Hypertens., vol. 37, no. 1, p. 1, Jul. 2018.
- G. S. Stergiou et al., “A Universal Standard for the Validation of Blood Pressure Measuring Devices,” Hypertension, vol. 71, no. 3, pp. 368–374, Mar. 2018.
- K. Bartels, S. A. Esper, and R. H. Thiele, “Blood Pressure Monitoring for the Anesthesiologist,” Anesth. Analg., vol. 122, no. 6, pp. 1866–1879, Jun. 2016.
- A. S. Vischer and T. Burkard, “Principles of Blood Pressure Measurement – Current Techniques, Office vs Ambulatory Blood Pressure Measurement.,” in Advances in Experimental Medicine and Biology, vol. 956, 2016, pp. 85–96.
- S. Yang, Y. Zhang, S. Y. Cho, R. Correia, and S. P. Morgan, “Non-invasive cuff-less blood pressure estimation using a hybrid deep learning model,” Opt. Quantum Electron., vol. 53, no. 2, pp. 1–20, 2021.
- M. Elgendi et al., “The use of photoplethysmography for assessing hypertension,” npj Digit. Med., vol. 2, no. 1, pp. 1–11, 2019.
- C. Wang et al., “Monitoring of the central blood pressure waveform via a conformal ultrasonic device,” Nat. Biomed. Eng., vol. 2, no. 9, pp. 687–695, Sep. 2018.
- C. Wang et al., “Continuous monitoring of deep-tissue haemodynamics with stretchable ultrasonic phased arrays,” Nat. Biomed. Eng., vol. 5, no. July, 2021.
- N. Luo et al., “Flexible Piezoresistive Sensor Patch Enabling Ultralow Power Cuffless Blood Pressure Measurement,” Adv. Funct. Mater., vol. 26, no. 8, pp. 1178–1187, 2016.
- J. Kim, E. F. Chou, J. Le, S. Wong, M. Chu, and M. Khine, “Soft Wearable Pressure Sensors for Beat-to-Beat Blood Pressure Monitoring,” Adv. Healthc. Mater., vol. 8, no. 13, pp. 1–9, 2019.
- D. Kireev et al., “Fabrication, characterization and applications of graphene electronic tattoos,” Nat. Protoc., vol. 16, no. 5, pp. 2395–2417, May 2021.
- “The Finapres® NOVA for continous non-invasive hemodynamics”. https://www.finapres.com/product/finapres-nova-basic/
- B. Ibrahim et al., “The Github Repository for Cuffless Blood Pressure Prediction Algorithms based on Graphene Bioimpedance Tattoos”. https://github.com/TAMU-ESP/Graphene_BP
Anyone can access the files, as long as they conform to the terms of the specified license.
License (for files):
Open Data Commons Attribution License v1.0
Total uncompressed size: 2.3 GB.
Access the files
- Download the ZIP file (751.2 MB)
- Download the files using your terminal:
wget -r -N -c -np https://physionet.org/files/bp-graphene-bioimpedance/1.0.0/
|data_trial37_bioz.csv (download)||5.5 MB||2022-03-11|
|data_trial37_finapresBP.csv (download)||452.1 KB||2022-03-11|
|data_trial37_finapresPPG.csv (download)||168.2 KB||2022-03-11|
|data_trial37_ppg.csv (download)||2.7 MB||2022-03-11|