Seizure Detection

This repository contains the winning submission for UPenn and Mayo Clinic's Seizure Detection Challenge on Kaggle.

It has been updated as of 18 Feb 2018 to include a requirements.txt for installing dependencies via pip. Small changes were necessary to get the repo working again with updated libraries. These changes were tested using sample_clip.mat from the Kaggle competition data rather than the original data which I did not have on hand. For the original submission without these changes, you may checkout the tag original-submission, e.g. git checkout original-submission.

http://www.kaggle.com/c/seizure-detection

This README and repository modelled on https://www.kaggle.com/wiki/ModelSubmissionBestPractices

Hardware / OS platform used

• 15" Retina MacBook Pro (Late 2013) 2.7GHz Core i7, 16GB RAM
• OS X Mavericks

Dependencies

Required

• python 2.7 with virtualenv
• hickle==2.1.0
• numpy==1.14.0
• scikit-learn==0.19.1
• scipy==1.0.0

Optional (to try out various data transforms)

• pywt (for Daubechies wavelet)
• scikits talkbox (for MFCC)

Installing dependencies

Setup the virtualenv and install the dependencies

virtualenv venv
. venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Train the model and make predictions

Activate the virtualenv

. venv/bin/activate

Obtain the competition data and place it in the root directory of the project.

seizure-data/
  Dog_1/
    Dog_1_ictal_segment_1.mat
    Dog_1_ictal_segment_2.mat
    ...
    Dog_1_interictal_segment_1.mat
    Dog_1_interictal_segment_2.mat
    ...
    Dog_1_test_segment_1.mat
    Dog_1_test_segment_2.mat
    ...

  Dog_2/
  ...

The directory name of the data should match the value in SETTINGS.json under the key competition-data-dir.

Then simply run:

python -m train

One classifier is trained for each patient, and dumped to the data-cache directory.

data-cache/classifier_Dog_1_fft-with-time-freq-corr-1-48-r400-usf-gen1_rf3000mss1Bfrs0.pickle
data-cache/classifier_Dog_2_fft-with-time-freq-corr-1-48-r400-usf-gen1_rf3000mss1Bfrs0.pickle
...
data-cache/classifier_Patient_8_fft-with-time-freq-corr-1-48-r400-usf-gen1_rf3000mss1Bfrs0.pickle

Although using these classifiers outside the scope of this project is not very straightforward.

More convenient is to run the predict script.

python -m predict

This will take at least 2 hours. Feel free to update the classifier's n_jobs parameter in seizure_detection.py.

A submission file will be created under the directory specified by the submission-dir key in SETTINGS.json (default submissions/).

Predictions are made using the test segments found in the competition data directory. They are iterated over starting from 1 counting upwards until no file is found.

i.e.

seizure-data/
  Dog_1/
    Dog_1_test_segment_1.mat
    Dog_1_test_segment_2.mat
    ...
    Dog_1_test_segment_3181.mat

To make predictions on a new dataset, simply replace these test segments with new ones. The files must numbered sequentially starting from 1 otherwise it will not find all of the files.

This project uses a custom task system which caches task results to disk using hickle format and falling back to pickle. First a task's output will be checked if it is in the data cache on disk, and if not the task will be executed and the data cached.

See seizure/tasks.py for the custom tasks defined for this project. More specifically the MakePredictionsTask depends on TrainClassifierTask, which means predict.py will train and dump the classifiers as well as make predictions.

Run cross-validation

python -m cross_validation

Cross-validation set is obtained by splitting on entire seizures. For example if there are 4 seizures, 3 seizures are used for training and 1 is used for cross-validation.

SETTINGS.json

{
  "competition-data-dir": "seizure-data",
  "data-cache-dir": "data-cache",
  "submission-dir": "./submissions"
}

• competition-data-dir: directory containing the downloaded competition data
• data-cache-dir: directory the task framework will store cached data
• submission-dir: directory submissions are written to

Model documentation

Available at https://github.com/MichaelHills/seizure-detection/raw/master/seizure-detection.pdf