Continuous high-density physiological signals from a surgical intensive care unit at Hospital Clínic de Barcelona

@article{PhysioNet-barcelona-sicu-signals-1.0.0,
  author = {{Marrero Garcia}, Ramses and {Cayuela Perez}, Rosalia and {Muñoz Rojas}, Guido and Carramiñana, Albert and Jacas, Adriana and Capdevila, Adriana and Alandes, Alexandre and {Vallverdú Martínez}, Jordi and Zattera, Luigi and Barbeta, Enric and Mercadal, Jordi and {Peñafiel Macías}, Petter Axcell and {Borrat Frigola}, Xavier and Ferrando, Carlos},
  title = {{Continuous high-density physiological signals from a surgical intensive care unit at Hospital Clínic de Barcelona}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0.0},
  doi = {10.13026/hgz9-s056},
  url = {https://doi.org/10.13026/hgz9-s056}
}

@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

Bedside monitors in intensive care units continuously generate high-resolution physiological signals that are rarely preserved systematically for research. This resource makes available to the community a set of multiparametric waveform recordings prospectively acquired in the surgical intensive care unit of Hospital Clínic de Barcelona (Spain). Data were captured with the open-source software VitalRecorder connected to Philips IntelliVue monitors, using a decentralized architecture without network transmission. Signals were recorded at their native sampling rates (electrocardiogram up to 500 Hz; plethysmography and invasive arterial pressure at 125 Hz; capnography, respiration and ventilator waveforms at 62.5 Hz; numeric trends at 1 Hz) and stored in the compressed binary .vital format. The set comprises 4,349 recordings (approximately 82 GB) totalling around 20,500 hours of monitoring over a twelve-month period. After privacy processing, 43 signal categories are published, all present in enough recordings to satisfy a k-anonymity threshold (k greater than or equal to 10, evaluated globally over the whole release, without temporal partitioning). The recordings, collected in pseudonymized form, are anonymized for this release: absolute timestamps, device and bed identifiers, patient data, and infrequent or potentially identifying signals are removed, leaving no re-identification key, and all times are relative to the start of each recording. Files are organized under a single folder with sequential names and are accompanied by per-recording metadata and documentation. The resource enables the development and validation of signal-based algorithms without the need for tabular clinical variables.

Background

Intensive care units continuously produce high-resolution physiological data, including electrocardiography, arterial pressure waveforms, respiratory signals, plethysmography and intracranial pressure, sampled between 1 Hz and 500 Hz, which dedicated acquisition systems can capture for both patient care and research [1]. This signal richness captures dynamic physiological processes that are not reflected in intermittently recorded clinical variables [1]. It also enables research avenues distinct from those supported by data derived from the electronic health record: biomedical signal processing, analysis of autonomic and cardiorespiratory variability [2], identification of signal-derived biomarkers, and development of predictive models [1]. Despite this potential, high-density physiological data are rarely preserved systematically [1], because institutional repositories tend to retain lower-density data: periodically charted vital signs, laboratory results, prescriptions and administrative data [3].

This gap is particularly visible in Europe. The openly available European ICU databases provide clinical data and time-aggregated parameters rather than raw waveforms [4]. Elsewhere, open tooling has made waveform-level resources feasible. VitalRecorder, a free tool for the automatic recording of high-fidelity data from multiple bedside devices [5], underpins VitalDB, a public multi-parameter vital signs database in surgical patients [6], and PhysioNet has long hosted complex physiological signal collections [7]. The surrounding ecosystem, which includes a browser-based viewer for high-resolution vital files [8] and a Python library for reading them [9], further lowers the barrier to reusing data in this format. The motivation for sharing this resource is twofold: to facilitate the development of signal-based models on a population and monitoring practice typical of a tertiary surgical unit, and to increase the diversity of settings represented in the open datasets available to the community.

Methods

Setting and acquisition

Data were acquired in the 13-bed surgical intensive care unit (UCIQ) of Hospital Clínic de Barcelona, a tertiary university hospital, serving a mixed population (postoperative, neurocritical, trauma and transplant). Prospective acquisition spanned twelve months, preceded by a preparation and pilot phase.

A decentralized architecture was adopted: the acquisition software ran locally on the existing bedside computers (Intel i5, 12 GB RAM, 256 GB SSD), connected directly to the monitor through a serial interface (RS232 to RJ45) with a custom pin configuration following the developer's specifications, storing the data locally without network transmission. Capture was performed with VitalRecorder [5], an open-source tool for real-time acquisition of biomedical signals, configured to start automatically in the shared generic sessions. Signals were recorded at their native rates and stored in the compressed binary .vital format [6]. A recording file corresponds to a continuous acquisition segment, typically delimited by a monitor reconnection, a sensor disconnection or a system restart.

Configured sampling rates:

Preparation, signal filtering and anonymization

Data were collected in pseudonymized form, and their anonymization was completed for this release through an automated procedure in Python (the vitaldb library [9]) that:

1. Recording merge: groups files by device, orders them chronologically, and merges into a single recording the consecutive files from the same device separated by less than 20 minutes (1,200 seconds).
2. Temporal and identifier anonymization: converts all timestamps into values relative to the start of each recording (first instant = 0), sets the time zone to UTC, and removes device and bed identifiers, so that no re-identification key remains.
3. Removal of auxiliary channels: discards demographic/administrative channels (the Demo/ prefix).
4. Naming harmonization (aliases): unifies equivalent names into 8 canonical names, so that clinically equivalent signals with device-dependent heterogeneous names (including arterial pressure measurements of invasive and non-invasive origin, and the different electrocardiogram leads) are published under a single label.
5. Signal exclusion: removes (a) infrequent or potentially identifying signals, on k-anonymity grounds — among them central venous pressure, pulmonary artery pressure, right atrial pressure, aortic pressure, left atrial pressure, skin/core temperature, train-of-four (TOF), anesthetic gases and events — and (b) derived or computed parameters that are not published in numeric form — such as the ST segment, cerebral perfusion pressure, cardiac indices and volumes — since they can be recomputed from the retained waveforms (see Usage Notes).
6. k-anonymity verification: over the whole set treated as a single block (without temporal partitioning), k is defined as the number of recordings containing each retained signal, and only signals with k above the threshold are kept. The minimum value is k = 11 (least prevalent signals: diastolic and systolic intracranial pressure).
7. Reorganization and documentation: recordings are published with sequential anonymous names (vitals_00001.vital, ...) under a single folder, accompanied by per-recording metadata and documentation.

The procedure preserves the original sampling rates, units and display ranges.

Signal unification map (alias to canonical):

Data Description

The set contains 4,349 recordings in the .vital format (VitalDB) [6], an open, compressed binary format optimized for physiological signals, readable with the open-source vitaldb library [9] and the free VitalRecorder software [1]. 43 signal categories are published (9 waveform and 34 numeric/derived). Each recording contains both waveform channels and numeric channels. Waveform samples are stored as 32-bit floats in physical units (e.g. mV for ECG, mmHg for pressures), so no gain/offset calibration is needed to read them.

File structure:

vitals/
    vitals_00001.vital
    vitals_00002.vital
    ...
RECORDS                      -> index listing all recordings
metadata.json                -> per-recording metadata (JSON)
k_anonymity_results.json     -> results of the k-anonymity analysis
CALIBRATION                  -> display scales for the waveform signals
README                       -> package documentation

Metadata: metadata.json contains one entry per recording with: session_id, filename, block_folder, duration_seconds, duration_minutes, n_tracks, n_waveform_tracks, n_numeric_tracks and file_size_mb. A top-level RECORDS file lists all recordings, and a CALIBRATION file gives the display scale (typical range and units) for each waveform signal.

Summary statistics:

• Recordings: 4,349

• Monitoring hours (total): approximately 20,500

• Published signal categories: 43 (9 waveform, 34 numeric/derived)

• Total size: approximately 82 GB

• Tracks per recording: median 13 (IQR 12–15)

• Duration per recording: median 92.5 min (IQR 30.2–296.5)

• Patients represented: an estimated 200–300 distinct patient episodes (no patient identifiers are retained, so an exact count is not available)

Inventory of published signals (number of recordings containing each signal and prevalence over the 4,348 valid recordings of the privacy analysis):

Usage Notes

As a newly released resource, the dataset has not yet been used in published downstream studies; its reuse potential and known limitations are outlined below.

How to read the files. With Python:

import vitaldb

# Load a record and list the available tracks
vf = vitaldb.VitalFile('vitals/vitals_00001.vital')
print(vf.get_track_names())

# Waveform tracks: resample at their native rate (ECG recorded at 500 Hz)
ecg = vf.to_pandas(['ECG_WAV'], interval=1/500, return_timestamp=True)

# Numeric tracks (e.g. HR): resample at 1-second intervals
# Numeric parameters are reported by the monitor at ~1 Hz; requesting them
# at the waveform rate would return a column that is mostly NaN.
numerics = vf.to_pandas(['HR'], interval=1, return_timestamp=True)

print(ecg.head())
print(numerics.head())

The files can also be opened with the free VitalRecorder software [5], or viewed in a web browser with the Vital File Web Viewer [8].

Reuse potential. The set is oriented toward research based solely on raw signal, without the need for tabular clinical variables: ECG processing, analysis of autonomic and cardiorespiratory variability, estimation of hemodynamic parameters from the arterial pressure waveform, neurological monitoring, and development of event detection or prediction models. The combination of 500 Hz ECG with invasive arterial pressure, capnography, intracranial pressure and EEG makes it suitable for several of these domains.

Derived variables not published. Certain parameters computed from the signal — such as the ST segment or cerebral perfusion pressure — are not published in numeric form. Because the waveforms from which they are derived are retained, these parameters can be recomputed from them; they were deliberately not derived at source so as not to introduce a-posteriori computed signals that, in the presence of motion artifacts or other interference, could yield erroneous measurements. Researchers who need these variables for a specific analysis, or to validate algorithms that compute them (for example, as a benchmark), may contact the authors.

Limitations users should be aware of

• Coverage is intermittent, not continuous: the set captures the unit's activity at intervals, not exhaustively.
• There is recording fragmentation: a relevant proportion of recordings is short.
• The availability of advanced signals is variable and reflects clinical practice (not a system failure). Mechanical ventilation waveforms depend on the ventilator model and its integration, so their availability should not be interpreted as the proportion of ventilated patients.
• Unified arterial pressures. The numeric arterial pressure measurements (BP_MEAN, SBP, DBP) unify invasive and non-invasive origin under a single label. The arterial pressure waveform (BP_WAV) is retained, so arterial-line values can be obtained from it; those who need the real origin of each measurement (for example, as a benchmark or for validation) may contact the authors.
• Unified ECG: the ECG_WAV waveform groups the different leads; the specific lead may vary between recordings.
• The set is anonymized and includes no link to clinical variables or patient identifiers; individual episodes cannot be reconstructed.
• Signals not included: central venous pressure, pulmonary artery pressure and other infrequent or sensitive signals are not part of the published set.

Complementary code and data. The file k_anonymity_results.json is included, with the detail of the privacy control (published signals, exclusions, aliases, k per signal and co-occurrence concordance). As complementary waveform datasets, VitalDB [6] and the waveform resources of the MIMIC family available on PhysioNet [7] may be of interest.

Release Notes

Version 1.0.0

Initial release

Ethics

The study protocol was reviewed and approved by the Research Ethics Committee for medicinal products (CEIm) of the Hospital Clínic de Barcelona (reference HCB/2026/0496) at its meeting of 30 April 2026; the committee complies with EMA/CHMP/ICH good clinical practice standards.

Acknowledgements

The authors thank the VitalLab team at Seoul National University Hospital for their sustained technical support and collaboration; the engineering students at the Universitat Politècnica de Catalunya for their contribution to the system deployment; and the unit's clinical staff for their sustained involvement in the data-capture protocols.

Conflicts of Interest

The author(s) have no conflicts of interest to declare.

References

1. De Georgia MA, Kaffashi F, Jacono FJ, Loparo KA (2015). "Information technology in critical care: review of monitoring and data acquisition systems for patient care and research". ScientificWorldJournal. 2015: 727694. doi: 10.1155/2015/727694.
2. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology (1996). "Heart rate variability: standards of measurement, physiological interpretation and clinical use". Circulation. 93(5): 1043–1065.
3. Johnson AEW, Bulgarelli L, Shen L, et al. (2023). "MIMIC-IV, a freely accessible electronic health record dataset". Sci Data. 10: 1.
4. Sauer CM, Dam TA, Celi LA, et al. (2022). "Systematic review and comparison of publicly available ICU data sets — a decision guide for clinicians and data scientists". Crit Care Med. 50(6): e581–e588. doi: 10.1097/CCM.0000000000005517.
5. Lee HC, Jung CW (2018). "VitalRecorder—a free research tool for automatic recording of high-fidelity clinical data in integrated operating room environments". Sci Rep. 8: 1688.
6. Lee HC, Park Y, Jang J, Jung CW (2022). "VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients". Sci Data. 9: 279.
7. Goldberger AL, Amaral LAN, Glass L, et al. (2000). "PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals". Circulation. 101: e215–e220.
8. Lee ER, Lee HC (2025). "Vital File Web Viewer: A Browser-Based Tool for Viewing High-Resolution Vital Signs Data" (version 1.0.0). PhysioNet. doi: 10.13026/3642-e205.
9. vitaldb: Python software for reading VitalDB [software]. PyPI; 2022. Available from: https://pypi.org/project/vitaldb/ [Accessed 10 September 2022].