• Introduction
• Getting Started
• Directory Structure
• TODO

The goal of this repo is to replicate and then improve upon the multi-class seizure classification problem described in this paper. The data we are using for this comes from the TUH EEG corpus online. It can be found here. The data directory is not sourced controled on Github, but see the Getting Started section for directions on how to download it.

The first thing to do is build out the data directory. Be sure to check the free space on your computer, as the raw data requires ~56 GB. To download the data run /nbs/00_get_data.ipynb. Then to build the data dictionary used for exploring the raw data set and building out the image data set, run the /nbs/01_data_dict.ipynb notebook.

• /nbs - all the notebook for exploring/building out the data pipeline
• /data
  • /raw - where the data got unzipped to. Follows TUH EEG format.
  • /augmented - where the sampled data was saved
    • /train_val
    • /test
  • /images - where image transforms of the
• /their_readme.txt - the readme from TUH EEG describing the data/eeg technology
• /data_dict.json - the data dictionary (generated in nbs) for ease of work/summary of the data
• /aug_labels.csv - a label file where each row is (path, label) for all files in the augmented directory
• /img_labels.csv - a label file where each row is (path, label) for all files in the images directory

• add percentage sampling for the script (to build small test cases)
• run the pipeline script for the small test set.
• Run the whole pipeline for the entire dataset
  • A little nervous about the size of the final directory (~21TB?)
• Worry about the validation split (by patient, session, etc?)
• Try out different techniques
  • Other down-sampling techniques
  • Alternatives to the fft (rfft, gramian angular field, wavelet)
  • Down sample less than 96