CAP:
Sleep classification across different machines

• Motivation: A recurrent problem in computational medicine is that models trained on data from a given recording device will not generalize to data coming from another device, even when both devices are from a similar equipment provider. Failure to generalize to unseen machines can cause critical issues for clinical practice because a false sense of confidence in a model could lead to a false diagnosis. We study these machinery-induced distribution shifts with the CAP dataset.
• Problem: We consider the sleep stage classification task from electroencephalographic (EEG) measurements. The dataset has five source domains, where each domain contains data gathered with a different machine. The goal is to generalize to unseen machines.
• Downloads: The preprocessed data is available on Academic Torrents and via Google Drive. It can also be directly downloaded through the WOODS dataset download script by running the following command:

  python -m woods.scripts.download_datasets CAP --data_path /path/to/data
  

• Preprocessing The preprocessing script can be found on the WOODS preprocessing script. You can run the preprocessing yourself by running the following command:

  python -m woods.scripts.fetch_and_preprocess CAP --data_path /path/to/data
  

References

[1] Terzano, Mario Giovanni, et al. “Atlas, rules, and recording techniques for the scoring of cyclic alternating pattern (CAP) in human sleep.” Sleep medicine 3.2 (2002): 187-199.

[2] Goldberger, Ary L., et al. “PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals.” circulation 101.23 (2000): e215-e220.

License

This dataset is licensed under the Open Data Commons Attribution license v1.0.