WildPPG: A Real-World PPG Dataset of Long Continuous Recordings (NeurIPS 2024, Official Code)

Manuel Meier, Berken Utku Demirel, Christian Holz

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

---

Quick Links:

• Project Website
• Paper
• Dataset

---

Loading the Dataset

The dataset is available for download here The dataset is split into .mat matlab files representing participants and can be loaded with matlab.

Loading the Data Using Python

The data can be loaded using scipy.io.loadmat. The function load_wildppg_participant(path) from src/heuristic_baselines/ppg_beats_analyzer.py can be used which cleans the loaded data and produces a nested dictionaries structure representing the same structure as the original .mat file.

Dataset Structure

Baseline Computation

For baseline computation, place all files of the dataset in a top-level data folder in this repository.

Heuristic Baselines

1. Add repo to path in matlab.
2. Execute src/heuristic_baselines/WildPPG_heuristic_baselines.m. This will produce temporary processing files containing PPG beat locations computed by the various methods.
3. python src/heuristic_baselines/ppg_beats_analyzer.py This will compare the derived heart rates to ecg-based ground truth, produce a pandas dataframe with various resulting measures, export it as a .csv file, and print the main results. The measures will be computed for every combination of participant, body_location, and method and include: MAE, MAPE, correlation, RMSE, as well as a measure for the fraction of windows in which the method was unable to compute HR values.

Supervised Baselines

1. Execute python src/supervised_baselines/data_preprocessing.py. This will produce a data file WildPPG.mat that serves as input-datafile for the supervised baslines

2. After the data is pre-processed and a input-datafile is available at src/supervised_baselines/data_preprocess/data, the command to run the supervised learning methods in src/supervised_baselines. Results are reported as print output.

python main_supervised_baseline.py --dataset 'alt' --backbone 'resnet' --block 8 --stride 2 --lr 5e-4 --batch_size 128 --n_epoch 999 --cuda 0

Citation

If your find our dataset, paper or codes useful, please cite our work:

@inproceedings{meier2024wildppg,
  title={WildPPG: A Real-World PPG Dataset of Long Continuous Recordings},
  author={Meier, Manuel and Demirel, Berken Utku and Holz, Christian},
  booktitle={The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2024},
}

License

All data is released under CC BY-NC-SA 4.0 license (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International)

Code of WildPPG is licensed under GNU General Public License version 3 (GPL-3.0)

Note: Some of the code to compute benchmarks is released under different licenses, please consult the license file for more detailed information.