snowballlab/Multisite-PPG

Multisite PPG Dataset

A multisite photoplethysmography (PPG) dataset: long-duration recordings from four body locations, synchronized activity logs, and ECG-derived heart-rate ground truth. It supports PPG-based HR estimation, signal-quality assessment, motion-artifact handling, and cross-site generalization research.

For data preprocessing and baseline training code, see our GitHub repository: wearable-ppg-dataset.

---

At a glance

Approx. size	~18.6 GB
Participants	20 (P1–P20)
Setting	Multi-day, free-living
Wearable sites	Earring, Necklace, Ring, Watch
Wearable modalities	Green PPG, IR PPG, 3-axis accelerometer, skin temperature
References	Polar H10 ECG + 3-axis accelerometer
HR ground truth	From synchronized Polar ECG (Pan–Tompkins + additional cleaning)

Sampling rates

Source	Signals	Rate
Wearable	PPG, accelerometer	100 Hz
Wearable	Temperature	1 Hz (interpolated to 100 Hz for alignment)
Polar	ECG	130 Hz
Polar	Accelerometer	25 Hz

---

Repository layout

Three main components ship in this repo:

Path	Role
raw_data/	Raw wearable streams, Polar references, activity logs
ppg_windowed_data/	8 s windows aligned with ECG-derived HR labels
sample_data/	Tiny subset for format checks / tutorials

Top-level tree:

Multisite-PPG/
├── raw_data/
├── ppg_windowed_data/
├── sample_data/
├── .gitattributes
└── croissant.json          # ML Croissant metadata

---

raw_data/

One folder per participant: wearable files at four sites, Polar references, and an activity log.

Wearable: P<N>_<Site>_raw.npz

<Site> ∈ {Earring, Necklace, Ring, Watch}.

Each file holds timestamped green + IR PPG, 3-axis accelerometer, skin temperature, and alignment metadata for Polar.

Polar references

File	Content
P<N>_polar_ecg_raw.npz	Polar H10 ECG
P<N>_polar_accl_raw.npz	Polar H10 3-axis accelerometer

Activity log

P<N>_activity_log.txt — timestamped labels as start <label> / end <label> pairs.

---

ppg_windowed_data/

Pre-windowed segments per participant and site, aligned with ECG-derived HR.

ppg_windowed_data/
├── P1/
│   ├── alignment_windows_P1_Earring.npz
│   ├── alignment_windows_P1_Necklace.npz
│   ├── alignment_windows_P1_Ring.npz
│   └── alignment_windows_P1_Watch.npz
├── P2/
│   └── ...
└── P<N>/
    └── ...

Each alignment_windows_*.npz includes: 8 s wearable PPG windows, motion channels, aligned ECG reference, ECG valid length, and HR labels.

Parameter	Value
Window length	8 s
Stride	1 s

---

sample_data/

Contains subsets of raw_data/ and ppg_windowed_data/, small enough for quick inspection without pulling the full release. Details in its own readme.

---

Quick start

Install

pip install huggingface_hub numpy

Download dataset

Sample only (~1.4 GB):

from huggingface_hub import snapshot_download
 
snapshot_download(
    repo_id="snowballlab/Multisite-PPG",
    repo_type="dataset",
    local_dir="Multisite-PPG",
    allow_patterns="sample_data/*",
)

Full dataset (~18.6 GB):

from huggingface_hub import snapshot_download
 
snapshot_download(
    repo_id="snowballlab/Multisite-PPG",
    repo_type="dataset",
    local_dir="Multisite-PPG",
)

---

Loading in Python

from pathlib import Path
import numpy as np

# Example file:
# Multisite-PPG/ppg_windowed_data/P1/alignment_windows_P1_Earring.npz
npz_path = Path("Multisite-PPG/ppg_windowed_data/P1/alignment_windows_P1_Earring.npz")

with np.load(npz_path) as z:  # allow_pickle=False by default
    ppg_green = z["ppg_green"]      # (N, T)
    ppg_ir = z["ppg_ir"]            # (N, T)
    hr_gt = z["hr_gt"]              # (N,)
    t0_ms = z["t0_ms"]              # (N,)
    t1_ms = z["t1_ms"]              # (N,)
    ppg_fs = float(z["ppg_fs"])     # scalar, typically 100.0
    ecg = z["ecg"]                  # (N, Lmax), zero-padded
    ecg_valid_len = z["ecg_valid_len"]  # (N,), valid ECG samples per window
    accel_x = z["accel_x"]          # (N, T)
    accel_y = z["accel_y"]          # (N, T)
    accel_z = z["accel_z"]          # (N, T)

print("ppg_green:", ppg_green.shape, "hr_gt:", hr_gt.shape, "ppg_fs:", ppg_fs)

# Example: recover valid (unpadded) ECG for one window
i = 0
ecg_i = ecg[i, : int(ecg_valid_len[i])]

• Use ppg_green or ppg_ir as model input windows.
• Use hr_gt as the heart-rate label aligned to each window.
• Use ecg_valid_len to remove right-side zero padding from ecg.

---

Intended uses

• PPG-based heart-rate estimation under motion
• Signal-quality assessment and artifact detection
• Cross-site generalization and multisite wearable modeling
• Benchmarking preprocessing and representation learning on biosignals

---

Limitations

• Free-living recordings: not strictly gap-free continuous sessions.
• Motion artifacts and low-quality segments occur naturally.
• Activity labels are user-logged and sparse, not exhaustive.
• Not validated for clinical decision support.

---

Ethics and anonymization

The study was approved by the institutional review board (IRB), and all participants provided informed consent for the public release of the de-identified dataset.

Participant identities are replaced with pseudonymous IDs (P1–P20), and direct personal identifiers are not included in the released dataset.

---

License

Released under CC BY-NC 4.0 (Attribution–NonCommercial).

Allowed: non-commercial research and education; derivative preprocessing with attribution.

Not allowed under CC BY-NC 4.0 without separate permission: commercial use.

Outside intended uses: re-identification attempts and clinical deployment without independent validation.

Full legal text: Creative Commons BY-NC 4.0.

---

Extra metadata

Machine-readable dataset description: see croissant.json at the repository root.

---