eth-siplab/WildPPG

WildPPG

Manuel Meier, Berken Utku Demirel, Christian Holz

Sensing, Interaction & Perception Lab, Department of Computer Science, ETH Zürich, Switzerland

---

📖 Overview

WildPPG provides 216 hours of continuous physiological recordings from 16 participants in diverse real-world scenarios.

It is the largest dataset for developing and evaluating wearable heart-rate detection algorithms.

---

• Participants, Duration & Ground truth for heart rate estimation:
  16 healthy adults; 810 min (~13.5 h) each, 216 h total. Ground truth heart rate: obtained from ECG with Pan–Tompkins and cleaned further during motions.

• Modalities & Placements:
  PPG, ECG (Lead I), 3-axis accel, skin temp, barometric altitude
  @ forehead · sternum · wrist · ankle

• Contexts & Environments:
  Mobility (walking, hiking, stairs), Stationary (meals, resting), Transit (car, train, cable car, elevator)
  Temperatures near 0 °C to full sun; altitudes up to 3 571 m

---

Quick Links: - Project Website - Paper

Loading the Dataset

The dataset is split into .mat MATLAB files representing participants and can be loaded with MATLAB.

Loading the Data

You can load the .mat file using either Python or MATLAB:

• Python:
  Use scipy.io.loadmat:

  from scipy.io import loadmat
  data = loadmat('WildPPG.mat')
  

• MATLAB:
  Use the built-in load function:

  data = load('WildPPG.mat');
  

The .mat file contains 14 cell arrays, each representing a variable (e.g., data_altitude_values, data_bpm_values, data_ppg_wrist).
Each cell array includes 16 entries, corresponding to data from 16 individual subjects.

Example loader

import numpy as np
import scipy.io

def load_domain_data(domain_idx):
    """Loads wrist PPG and heart rate data for a single subject (domain).

    Args:
        domain_idx (int): Index of the subject (0–15).

    Returns:
        X (np.ndarray): PPG signal data (n_samples × signal_dim).
        y (np.ndarray): Heart rate values (bpm), adjusted to start from 0.
        d (np.ndarray): Domain labels (same shape as y), equal to domain_idx.
    """
    data_path = 'data/WildPPG/WildPPG.mat'
    data_all = scipy.io.loadmat(data_path)

    # Load PPG signal and heart rate values
    data = data_all['data_ppg_wrist']
    data_labels = data_all['data_bpm_values']

    domain_idx = int(domain_idx)
    X = data[domain_idx, 0]
    y = np.squeeze(data_labels[domain_idx][0]).astype(int)

    # Mask out invalid samples (e.g., NaNs, infs, and HR < 30 bpm)
    mask_Y = y >= 30
    mask_X = ~np.isnan(X).any(axis=1) & ~np.isinf(X).any(axis=1)
    combined_mask = mask_Y & mask_X

    X = X[combined_mask]
    y = y[combined_mask] - 30  # Normalize HR: min HR is 30 bpm → range starts from 0
    d = np.full(y.shape, domain_idx, dtype=int)

    return X, y, d

Load PPG & Heart Rate Data for a Single Subject

The function load_domain_data(domain_idx) loads wrist PPG signal data and heart rate (HR) labels for a single subject from the WildPPG.mat file.

• domain_idx ranges from 0 to 15, each corresponding to one of the 16 subjects.
• The data is preprocessed by:
  • Removing invalid samples (NaNs, infs, and HR < 30 bpm)
  • Normalizing HR values to start from 0 (by subtracting 30 bpm)
• The function returns:
  • X: preprocessed PPG signal (shape: n_samples × signal_dim)
  • y: adjusted heart rate labels
  • d: domain label (same shape as y, filled with domain_idx)

Example usage:

x, y, d = load_domain_data(3)  # Load data for subject 3