# Continuous high-density physiological signals from a surgical intensive care unit at Hospital Clínic de Barcelona
## Description
This dataset contains continuous, multiparametric physiological waveform recordings
obtained from **Philips IntelliVue** bedside monitors in the surgical intensive care
unit (UCIQ) of **Hospital Clínic de Barcelona** (Spain). Data were captured with the
open-source **VitalRecorder** software and stored in the `.vital` format (VitalDB).
The recordings were collected in pseudonymized form and their anonymization was
completed for this release:
- No absolute dates or times (all timestamps are relative to the start of each recording)
- No device or bed identifiers
- No patient information (no re-identification key remains)
- Time zone set to UTC
- Infrequent or potentially identifying signals removed (k-anonymity control, k >= 10)
## Dataset statistics
| Metric | Value |
| --------------------------- | ----------------------------------- |
| Recordings | 4,349 |
| Total monitoring time | approx. 20,479 hours |
| Total size | 82,112 MB (approx. 80 GB) |
| Total tracks | 63,763 |
| Waveform tracks | 16,326 |
| Numeric tracks | 47,437 |
| Published signal categories | 43 (9 waveform, 34 numeric/derived) |
## File structure
```
vitals/
vitals_00001.vital
vitals_00002.vital
...
RECORDS -> index listing all recordings (one path per line)
metadata.json -> per-recording metadata
k_anonymity_results.json -> results of the k-anonymity analysis
CALIBRATION -> display scales (range and units) for the waveform signals
README.md -> this file
```
`metadata.json` provides 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`.
## File format
The files are in the `.vital` format (VitalDB), an open, compressed binary format
optimized for physiological signals. Each file contains both waveform and numeric
tracks. Signals were recorded at their native sampling rates (ECG up to 500 Hz;
plethysmography, invasive arterial pressure and intracranial pressure at 125 Hz;
capnography, respiration and ventilator waveforms at 62.5 Hz; numeric trends at 1 Hz).
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.
### Waveform tracks (9)
`ECG_WAV` (electrocardiogram), `PLETH` (plethysmography), `BP_WAV` (arterial pressure
waveform), `CO2` (capnography), `RESP` (respiration), `EEG` (electroencephalogram),
`AWP_WAV` (airway pressure), `FLOW_WAV` (airway flow), `ICP` (intracranial pressure).
### Numeric tracks (selected)
`HR` (heart rate), `PLETH_SAT_O2` (oxygen saturation, SpO2), `SBP`, `DBP`, `BP_MEAN`
(arterial pressures), `RR` (respiratory rate), ventilator parameters
(`VENT_RR`, `TV_EXP`, `MV_EXP`, `PEEP_CMH2O`, `PIP_CMH2O`, `FIO2`, ...),
capnography numerics (`AWAY_CO2_ET`, ...), `EEG_BIS` and related indices,
`ICP_MEAN`/`ICP_SYS`/`ICP_DIA`, and other clinical parameters. The full inventory of
the 43 published signals, with per-signal prevalence, is in `k_anonymity_results.json`.
### Signal harmonization (aliases)
Clinically equivalent signals with device-dependent names are published under a single
canonical name:
- `BP_MEAN`, `SBP`, `DBP`: unified from invasive (ART_*, ABP_*) and non-invasive (NIBP_*) sources
- `HR`: unified from ECG_HR, NIBP_HR, ABP_HR
- `BP_WAV`: unified from ART, ART_WAV, ABP
- `ECG_WAV`: unified from all ECG leads (ECG_I/II/III/V and their _WAV variants)
- `CO2`: unified from ETCO2, RR_CO2
Oxygen saturation is published under its native name `PLETH_SAT_O2`.
### Signals not published
Removed on k-anonymity grounds: central venous pressure (CVP), pulmonary artery
pressure (PAP), right atrial pressure (RAP), aortic pressure (AOP), left atrial pressure
(LAP), skin/core temperature, train-of-four (TOF), anesthetic gases (GAS1_*) and events.
Removed as derived/computed parameters (not published in numeric form): ST segment,
cerebral perfusion pressure, cardiac indices (CI), and volumetric parameters (VOL_*).
These can be recomputed from the retained waveforms; they were deliberately not derived
at source to avoid introducing computed signals that, under 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 (e.g. as a
benchmark) may contact the authors.
Demo/test channels (`Demo/`) were also removed.
## How to read the files
With Python (vitaldb library):
```python
import vitaldb
vf = vitaldb.VitalFile('vitals/vitals_00001.vital')
print(vf.get_track_names())
df = vf.to_pandas(['ECG_WAV', 'HR'], interval=1/500, return_datetime=False)
```
The files can also be opened with the free VitalRecorder software (vitaldb.net).
## k-anonymity analysis
A k-anonymity control was performed over the whole release as a single block (no
temporal partitioning), with k defined as the number of recordings containing each
retained signal:
- Global k = 11 (limiting signals: ICP_DIA, ICP_SYS)
- 43 published signals after exclusion and alias unification
- A co-occurrence (Jaccard) concordance report is included in `k_anonymity_results.json`
## Ethics
The study protocol was approved by the Research Ethics Committee for medicinal products
(CEIm) of Hospital Clínic de Barcelona (reference HCB/2026/0496) on 30 April 2026. The
sponsor is the Fundació de Recerca Clínic Barcelona – IDIBAPS.
## License
Creative Commons Attribution 4.0 International (CC BY 4.0). The data may be freely used,
shared and adapted, provided appropriate credit is given.