Current work on seizure prediction focuses on classifying segments at any given time into "seizure" or "not seizure". Given that solutions to this problem are rapidly improving, we ask another question: how long until a seizure occurs?
Even the best model's predictions were poorly correlated with results (R^2 = 0.242). However, we expect that some of the code will remain useful. You can see more in our paper. Beware that the code is research-quality; if something is broken, feel free to email us and we'll try to fix it. We would also appreciate a heads-up if you use our work.
More precisely, we currently frame the question as "given that there will be a seizure soon, when will it be?" We use the CHB-MIT dataset to ask that question. We split the recording up into five-second segments, then discard all segments that either:
- Overlap with a seizure (as this would incentivize the classifier to learn ictal patterns over all else)
- Has an indeterminate time until the next seizure (i.e. occurs after the last recorded seizure in a given file)
Note that the second requirement leads to substantial right-censorship. In the future, we'll explore techniques from survival analysis to mitigate this issue.
We're currently following Tsiouris et al.
(2018) for our features.
Refer to that paper for details on what each of the features mean. The order of
the features in their vectors is documented in
references/feature-descriptions.txt
This repository currently contains a variety of models that attempt to predict based solely on the features in a single model:
- Dummy (predicts the mean, only used as a baseline)
- Linear (ordinary least squares, no transformations)
- Multi-layer perceptron (with hidden layer sizes 60 and 10)
- Random forest (with 100 trees)
- Gradient boosted trees (all XGBoost defaults)
Due to the poor performance of all of these (see below), we intend to move on to more powerful methods. Our next step will be using a convolutional network to examine multiple adjacent five-second segments. We may also use an LSTM if the convolutional network does not perform well.
| Model | R^2 | Adjusted R^2 |
|---|---|---|
| Dummy | -0.184 | -0.231 |
| Linear | +0.128 | +0.093 |
| MLP | -0.608 | -0.672 |
| XGBoost | -0.499 | -0.558 |
| Convolutional (MSE) | -0.013 | |
| Convolutional (Custom) | +0.105 | |
| LSTM (Custom) | +0.242 |
See the reports/figures directory for predicted-actual plots.
- Install NumPy, then the requirements from the
requirements.txtfile. This is because pyEDFlib will refuse to install unless NumPy is already present. - Run
create-directories.sh - Download the data from the CHB-MIT dataset with the
fetch-data.shscript (requireswget) - Calculate the feature values. This can be done with the
build_features.pyscript insrc/features.pyfor raw features, thennormalize_and_split.pyfor scaling them for use in a neural network or other non-scale invariant model - Run the individual models in the
modelsdirectory. They are all split up into training and predicting steps.
This code has all been written by Jeremy Angel and Matthew Wootten, except where comments explicitly indicate otherwise. It is licensed under the GPLv3.
To cite this repository, you can use the following:
@misc{AngelWootten_2019,
author = {Jeremy Angel and Matthew Wootten},
title = {Convolutional neural networks for real-time seizure forecasting},
year = 2019,
url = {https://github.com/matthewlw/seizure-forecasting}
}