GlucoseDao/ppg-windows

PPG windows

Finger and wrist photoplethysmography (PPG) cut into quality-checked 30-second windows at 25 and 125 Hz, ready to train on, with each person's age, sex and, for VitalDB, outcome flags. It is the training data of kron-ppg, a streaming KronLSTM PPG model whose first task is an age clock, published so others can skip downloading about 18 GB of raw tracks and preparing them. 25 Hz is the rate of phones and watches (and kron-ppg); 125 Hz is the rate of MIMIC/PhysioNet finger PPG and of models such as PaPaGei. A source is never upsampled: GalaxyPPG, recorded at 25 Hz, has only 25 Hz.

config	windows	people	train / validation / test people
vitaldb_25hz	1,195,506	5,887	4,071 / 912 / 897
vitaldb_125hz	1,195,923	5,887	4,071 / 912 / 897
bidmc_25hz	820	53	33 / 14 / 6
bidmc_125hz	821	53	33 / 14 / 6
galaxyppg_25hz	1,847	24	15 / 5 / 3

Each source also has a <source>_subjects config: one row per person, with the same split column.

Columns

Windows (<source>_<rate>hz configs):

column	type	meaning
subject	string	the person, as the source names them (VitalDB subjectid)
recording	string	the recording the window was cut from (VitalDB caseid)
start_s	float64	where the window starts in its recording, seconds
hr_bpm	float32	heart rate estimated from the window's autocorrelation
periodicity	float32	the autocorrelation at that beat period (the quality gate keeps >= 0.35)
signal	float16[30 x rate]	the window, robust-scaled (median 0, IQR/1.349 = 1): 750 samples at 25 Hz, 3750 samples at 125 Hz

vitaldb_25hz, vitaldb_125hz also carry the vitals measured over each window, read from VitalDB's other tracks (monitor numerics and ECG lead II, 500 Hz), joined on (recording, start_s) from the 25 Hz windows. A value is null where its track is missing, out of a plausible range, or (ECG, pulse arrival time) too few clean beats were found; a window with no 25 Hz twin has none. Windows with at least one vital: vitaldb_25hz 100%, vitaldb_125hz 96%.

column	type	meaning
hr	float64	monitor heart rate, bpm (median over the window)
spo2	float64	monitor oxygen saturation, %
bt	float64	body (core) temperature, degC
nibp_sbp	float64	cuff systolic pressure, mmHg (a reading is held until the next, every few minutes)
nibp_dbp	float64	cuff diastolic pressure, mmHg
nibp_mbp	float64	cuff mean pressure, mmHg
art_sbp	float64	arterial-line systolic pressure, mmHg (beat by beat; about half the cases)
art_dbp	float64	arterial-line diastolic pressure, mmHg
art_mbp	float64	arterial-line mean pressure, mmHg
sbp	float64	systolic pressure: the arterial line's where there is one, else the cuff's
dbp	float64	diastolic pressure, chosen the same way
ecg_hr	float64	heart rate from ECG lead II R peaks, bpm
rr_ms	float64	median R-R interval, ms
sdnn_ms	float64	SDNN over the window's R-R intervals, ms
rmssd_ms	float64	RMSSD over the window's R-R intervals, ms
pat_ms	float64	pulse arrival time: R peak to the steepest PLETH upstroke before the next beat, ms (median). VitalDB's PLETH lags its ECG by a device delay, so values run 0.6-0.8 s: compare within a person and monitor, not with published PATs
beats	float64	clean beats the ECG features rest on
bp_source	string	art, nibp, or empty: where sbp/dbp came from

Subjects (<source>_subjects configs): subject, age (years, null when unknown), sex, the source's other per-person fields, and split. VitalDB adds caseid, asa, preop_htn, preop_dm, death_inhosp, department, optype, emop and bmi from its case table.

How the windows were made (changes to the sources)

For each recording, at each rate: fill gaps of up to 0.5 s; band-pass 0.5-8 Hz (4th-order Butterworth, zero phase); resample to the rate (polyphase); cut 30 s windows without overlap; scale each window robustly. A window is kept when it is finite, not flat, not clipped and its autocorrelation at a heart rate of 30-200 bpm is at least 0.35. At most 240 windows are kept per recording, spread evenly over it, so a three-hour surgery does not outweigh an eight-minute ICU record. Each rate is gated on its own, so the window sets of two rates overlap closely but not exactly; join them on (recording, start_s).

The split is by person, from a hash of <source>/<subject>, so a person's windows are never on two sides, every rate has the same split, and the split does not move when people are added.

Loading

import polars as pl
train = pl.read_parquet("hf://datasets/GlucoseDao/ppg-windows/vitaldb/25hz/train-*.parquet")
x = train["signal"].to_numpy()          # (n, 750) float16

or datasets.load_dataset("GlucoseDao/ppg-windows", "vitaldb_25hz", split="train").

What it is not

Age from PPG is a calendar-age regression target. A predicted age, or the gap between it and the real age, is not a biological age or a diagnosis. VitalDB people were having surgery and BIDMC people were in intensive care, so neither is a healthy population.

Sources, licences and credit

vitaldb (CC BY 4.0, https://creativecommons.org/licenses/by/4.0/)

VitalDB: finger PPG (SNUADC/PLETH, 500 Hz) recorded during surgery at Seoul National University Hospital, with age, sex, ASA class, preop_htn, preop_dm, death_inhosp, department, operation type, emergency flag and BMI per case. One case per person (the first by case id, of those with a PLETH track and an age). From the VitalDB open API (https://vitaldb.net/dataset/), the same data as https://physionet.org/content/vitaldb/1.0.0/.

Credit to VitalDB is required (citation above). The vitaldb.net data use agreement also asks users not to attempt to identify any person from the data.

Lee HC, Park Y, Yoon SB, Yang SM, Park D, Jung CW. VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients. Sci Data. 2022;9(1):279. doi:10.1038/s41597-022-01411-5. Data: Lee HC, Jung CW. VitalDB (version 1.0.0). PhysioNet. doi:10.13026/czw8-9p62.

bidmc (ODC-By 1.0, https://opendatacommons.org/licenses/by/1-0/)

BIDMC PPG and Respiration: 53 eight-minute ICU recordings of finger PPG at 125 Hz with age and sex, from the MIMIC-II matched waveform database. Age 90+ is stored as 90 with age_censored = 1. From https://physionet.org/content/bidmc/1.0.0/.

Contains information from the BIDMC PPG and Respiration Dataset, which is made available under the ODC Attribution License.

Pimentel MAF et al. Toward a robust estimation of respiratory rate from pulse oximeters. IEEE Trans Biomed Eng. 2017;64(8):1914-1923. doi:10.1109/TBME.2016.2613124. Goldberger AL et al. PhysioBank, PhysioToolkit, and PhysioNet. Circulation. 2000;101(23):e215-e220.

galaxyppg (CC BY 4.0, https://creativecommons.org/licenses/by/4.0/)

GalaxyPPG: Samsung Galaxy Watch5 green PPG at 25 Hz from 24 people aged 20-29, worn through a protocol of rest and activity. The watch's signal is inverted so the systolic peak points up, as on a finger sensor. From https://zenodo.org/records/14635823.

Credit to the GalaxyPPG authors is required (citation above).

Park, Zheng & Lee (2025). GalaxyPPG: a multi-device PPG dataset [data set]. Zenodo. doi:10.5281/zenodo.14635823.

This dataset adds no terms of its own: each source's rows are under that source's licence, with the changes listed above.