PulseDB-Ext: Blood Pressure and Synthetic Modeled Quantities

@article{PhysioNet-pulsedb-ext-synthetic-model-1.0.0,
  author = {Belik, Filip and Crandall, Henry and Schuessler, Tyler and Hohenegger, Christel and Osting, Braxton and Sanchez, Benjamin},
  title = {{PulseDB-Ext: Blood Pressure and Synthetic Modeled Quantities}},
  journal = {{PhysioNet}},
  year = {2026},
  month = nov,
  note = {Version 1.0.0},
  doi = {10.13026/zqga-jh51},
  url = {https://doi.org/10.13026/zqga-jh51}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

Contained in the database are time-series-split, Fourier truncated blood pressure signals scraped from the PulseDB database along with corresponding modeled time-series of interest. Each entry contains a period length and 100 uniformly spaced values in time. The modeled time-series of interest include the arterial radii, the bulk blood conductivity, and the bioimpedance measured at the wrist. This database is provided to support the development, reproducibility, and benchmarking of physics-informed machine-learning methods for cuffless hemodynamic monitoring.

Background

Continuous, cuffless blood-pressure (BP) monitoring could improve hypertension management by enabling frequent measurements during daily life, whereas conventional cuff-based devices provide only intermittent measurements. However, many wearable cuffless BP approaches rely on pulse-wave analysis (PWA) of photoplethysmography (PPG) or tonometry waveforms, with or without pulse-arrival-time features. In the Microsoft Research Aurora Project, these approaches provided essentially no additional value over a baseline model without waveform measurements for estimating resting or 24-hour ambulatory BP [1]. A separate Microsoft Research analysis likewise found that PPG-to-BP estimation is an ill-conditioned problem: similar PPG waveforms may correspond to meaningfully different BP waveforms [2].

A complementary approach is to measure a signal with a defined biophysical relationship to arterial hemodynamics. Pulsatile blood flow and pressure cause cyclic arterial-wall motion; consequently, arterial radius varies throughout the cardiac cycle [3]. Electrical bioimpedance (BioZ) measurements are sensitive to this change in arterial blood volume and to the electrical conductivity of blood. In a wrist measurement, arterial radius and bulk blood conductivity therefore serve as intermediate physiological states that connect pressure-driven flow to the electrical resistance measured at the skin surface. Independent groups have demonstrated cuffless BP estimation with wearable BioZ sensors, including a ring-based system and graphene bioimpedance tattoos [4,5].

The present dataset enables development and evaluation of models that exploit this mechanistic pathway. Directly collecting synchronized wrist BioZ, invasive arterial BP, arterial geometry, and blood electrical-property waveforms from a large and diverse cohort is impractical. We therefore use real radial arterial BP waveforms from PulseDB [6] as physiologically realistic inputs to a multiscale forward model. The model uses fluid dynamics and arterial-wall mechanics to generate time-varying arterial radius and blood conductivity, then uses an electrical model to generate the corresponding wrist-surface BioZ waveform. This produces 100,000 paired BP–BioZ examples, along with the intermediate physiological time series, for training and benchmarking physics-informed machine-learning models. The synthetic signals provide known ground truth for the otherwise unobservable states and permit controlled evaluation of the inverse task: recovering BP and hemodynamic quantities from a wearable BioZ signal. This hybrid use of experimental BP waveforms and synthetic BioZ signals follows the modeling and inverse-problem validation framework described in [7].

Methods

Synthetic training data were sourced from pressure waveforms obtained from PulseDB [6], a publicly available waveform database. Individual cardiac cycles were extracted from the continuous recordings by identifying local minima in the blood pressure signal and segmenting the photoplethysmography and blood pressure waveforms at each successive minimum, yielding 1,261,832 single-period packets spanning one complete cardiac cycle. No additional preprocessing or filtering was applied to the waveforms prior to segmentation. After segmentation, the signals were cast onto 100 uniform time points between 0 and the corresponding period using cubic spline interpolation. Then, 100,000 of the periodic signals were randomly subselected, without replacement, subject to the following conditions: (a) the heart rate was between 40 and 200bpm, (b) the blood pressure was not below 90/60 (hypotension), and (c) the blood pressure was not above 140/90 (stage 2 hypertension). Signals were made periodic through a truncation of the zeroth through tenth order Fourier modes and shifted to be min-to-min in blood pressure. Synthetic time-series beyond the blood pressure waveforms were generated using a two-component biophysical forward model detailed in the Appendix of the associated publication [7].

The first component is a hemodynamic model based on the linearized Navier–Stokes equations in an elastic tube, following the framework of Zamir (2000) [3]. This model couples pulsatile blood flow in a branched arterial tree to elastic vessel-wall motion, yielding the time-varying luminal radius $a(t)$ of the radial artery over each cardiac cycle. Red blood cell (RBC) orientation and deformation under shear flow are incorporated via Maxwell–Fricke mixture theory, from which the bulk electrical conductivity of blood $\sigma_{bl}(t)$ is derived as a function of instantaneous wall shear stress.

The second component is a three-layer cylindrical bioimpedance model that translates the computed blood conductivity into a surface-measurable wrist impedance signal $\text{Re}Z$. The model treats the wrist cross-section as concentric cylindrical layers — skin, subcutaneous adipose tissue, and muscle surrounding the radial artery — and solves a reduced form of Maxwell's equations to compute the total real-part bioimpedance as seen by surface electrodes.

For each of the 100,000 cardiac cycles, the seven biophysical parameters ($a$, $\sigma_1$, $H$, $\text{hr}$, $d_2$, $E$, $\bar{\tau}_w$) were drawn from physiologically plausible distributions documented in Supplementary Table 1 of the associated publication [7]. The Biolvar folders provide additional data in which each parameter is varied individually while others are held at their nominal values, enabling sensitivity analysis and model calibration experiments. Full mathematical details, including the governing equations and parameter distributions, are provided in the Appendix of the associated publication [7].

Data Description

The file BP_min2min.csv contains all 100,000 blood pressure signals along with anonymized data including the subject ID, age, and gender. The file a_min2min.csv contains the modeled radial arterial radius as a function of time. The file SBl_min2min.csv contains the modeled bulk radial arterial conductivity as a function of time. The file ReZ_min2min.csv contains the modeled bioimpedance (BioZ) measured at the wrist as a function of time. In each Biolvar folder, there is a file containing randomly sampled metaparameters along with corresponding BioZ waveforms. The Biolvar folders correspond to the following parameters: E for arterial wall Young's Modulus, H for hematocrit level, a for resting arterial radius, d2 for fat layer thickness, hr for heart rate, mean_tau_w for mean wall shear stress, and sigma_1 for conductivity of the skin. In each Biolvar folder, parameters are individually randomly sampled from parameter distributions detailed in [7] while all others are held constant. In Biolvar_all/, all parameters are randomly sampled at the same time and the data is split into 5 parts.

BP_min2min.csv — Blood pressure waveforms.

a_min2min.csv — Modeled radial arterial radius.

SBl_min2min.csv — Modeled bulk blood conductivity.

ReZ_min2min.csv — Modeled wrist bioimpedance.

Biolvar_*/metaparams.csv — Biophysical model parameters.

Biolvar_*/ReZ_min2min.csv — Bioimpedance waveforms paired with the above parameter sets. Columns follow the same SubjectID + 0 … 99 structure as ReZ_min2min.csv. In Biolvar_all/, all parameters are co-varied simultaneously; data are split into five equal parts.

Usage Notes

This dataset was originally used to train and evaluate the signal-tagged physics-informed neural network (sPINN) described in "Cuffless hemodynamic monitoring with physics-informed machine learning models" [7], where it served as the synthetic pre-training corpus enabling population-level and subject-specific cuffless blood pressure prediction from wrist bioimpedance signals.

Reuse potential. Beyond the original application, this dataset is well suited for: (i) training or fine-tuning alternative machine learning architectures (e.g., convolutional networks, transformers, or physics-informed variants) for cuffless BP regression; (ii) calibrating or validating other biophysical forward models of arterial hemodynamics or wrist bioimpedance against a large, physiologically diverse set of waveforms; (iii) benchmarking the sensitivity of model predictions to individual biophysical parameters using the Biolvar files, which isolate each parameter's effect in a controlled fashion; and (iv) developing and testing signal-processing pipelines for periodic cardiovascular waveforms.

Known limitations. The blood pressure waveforms are restricted to normotensive and mildly hypertensive ranges reflecting the source distribution in PulseDB; performance of models trained on this data for hypertensive or hypotensive populations should be interpreted with caution. The synthetic modeled quantities (arterial radius, conductivity, bioimpedance) are outputs of a forward model with assumed parameter distributions and a simplified cylindrical wrist geometry; they do not constitute direct physiological measurements. Subject metadata (age, sex) are drawn from PulseDB and reflect its demographic composition.

Complementary resources. The PulseDB database [6] is the source of all blood pressure waveforms and may be of interest for researchers seeking additional waveform types (e.g., photoplethysmography).

Release Notes

Version 1.0.0

Initial release

Ethics

The blood pressure waveforms used are available in the open-source database, PulseDB. The remaining data is generated with the forward model detailed in "Cuffless, calibration-free hemodynamic monitoring with physics-informed machine learning models".

Conflicts of Interest

Dr. Sanchez is a co-founder of and holds equity in Haystack Diagnostics, Inc.
He holds equity and serves as Scientific Advisor to B-Secur, Ltd., and Sobr Safe, Inc.
He holds equity and serves as a Chief Scientific Officer of Hemodynamiq, Inc.
He serves as a Chief Scientific Advisor to First Capital Ventures, LLC, and Chief Scientific Officer to Promptus, LLC. He holds equity and serves as Head of Biosensing and Product Development of NeuralPoint AI, Inc. The other authors have no conflicts of interest to declare.

References

1. Mukkamala R, Shroff SG, Landry C, Kyriakoulis KG, Avolio AP, Stergiou GS. The Microsoft Research Aurora Project: important findings on cuffless blood pressure measurement. Hypertension. 2023;80(3):534-540. doi:10.1161/HYPERTENSIONAHA.122.20410.
2. Mehta S, Kwatra N, Jain M, McDuff D. Examining the challenges of blood pressure estimation via photoplethysmogram. Sci Rep. 2024;14:18318. doi:10.1038/s41598-024-68862-1.
3. Zamir M. The physics of pulsatile flow. New York (NY): Springer; 2000.
4. Sel K, Osman D, Huerta N, Edgar A, Pettigrew RI, Jafari R. Continuous cuffless blood pressure monitoring with a wearable ring bioimpedance device. NPJ Digit Med. 2023;6(1):59. doi:10.1038/s41746-023-00796-w.
5. Kireev D, Ibrahim B, Kumar N, Akbari A, Akinwande D, Jafari R. Continuous cuffless monitoring of arterial blood pressure via graphene bioimpedance tattoos. Nat Nanotechnol. 2022;17:864-870. doi:10.1038/s41565-022-01145-w.
6. Wang W, Mohseni P, Kilgore KL, Najafizadeh L. PulseDB: a large, cleaned dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood pressure estimation methods. Front Digit Health. 2023;4:1090854. doi:10.3389/fdgth.2022.1090854.
7. Crandall H, Schuessler T, Bělík F, et al. Cuffless hemodynamic monitoring with physics-informed machine learning models. Nat Commun. 2026;17:6429. doi:10.1038/s41467-026-72693-1.

Cited by

4 publications cite this project. The most recent are listed below.

• On the Mechanics of Bioimpedance Signals Under Deformation
  Chihtong Lee, William Chan Santos, Carmina Andrade et al. Journal of Applied Mechanics, 2026.
• System-level design of flexible pressure sensors for arterial pulse monitoring: Waveform fidelity from materials to algorithms
  Zeying Zhang, Letian Hao, Bokun Zhang et al. Chemical Engineering Journal, 2026.
• Toward Noninvasive Calibration-Free Blood Pressure Estimation: A Physics-Informed Neural Network With Hemodynamic Priors
  Federico Delrio, Vincenzo Randazzo, Giansalvo Cirrincione et al. IEEE Sensors Journal, 2026.
• Methodologies and Techniques for Metrological Characterization of Wearable Devices
  Chen Liu, Leonardo Bianchi, Paola Saccomandi. IEEE Sensors Reviews, 2026.

View all 4 publications

Citation data from OpenAlex. Papers that use the project without citing it may be missing.