Siyi Tang, Jared A. Dunnmon, Khaled Saab, Xuan Zhang, Qianying Huang, Florian Dubost, Daniel L. Rubin, Christopher Lee-Messer, arXiv, https://arxiv.org/abs/2104.08336

Automated seizure detection and classification from electroencephalography (EEG) can greatly improve the diagnosis and treatment of seizures. While prior studies mainly used convolutional neural networks (CNNs) that assume image-like structure in EEG signals or spectrograms, this modeling choice does not reflect the natural geometry of or connectivity between EEG electrodes. In this study, we propose modeling EEGs as graphs and present a graph neural network for automated seizure detection and classification. In addition, we leverage unlabeled EEG data using a self-supervised pre-training strategy. In summary, our graph-based modeling approach integrates domain knowledge about EEG, sets a new state-of-the-art for seizure detection and classification on a large public dataset, and provides better ability to identify seizure regions.

We use the Temple University Seizure Corpus (TUSZ) v1.5.2 in this study. The TUSZ dataset is publicly available here. After you have registered and downloaded the data, you will see a subdirectory called edf which contains all the EEG signals and their associated labels. We use the EEG files in the edf/dev subfolder as our held-out test set. We further split the EEG files in the edf/train subfolder into train and validation sets by patients. See folders ./data/file_markers_detection, ./data/file_markers_classification, and ./data/file_markers_ssl for details.

In this study, we exclude five patients from the test set who exist in both the official TUSZ train and test sets. You can find the list of excluded patients' IDs in ./data_tusz/excluded_test_patients.txt.

In addition, ./data_tusz/focal_labeled_as_generalized.csv provides the list of 27 seizures in the test set that we think are focal seizures (manually analyzed by a board-certified EEG reader) but are labeled as generalized non-specific seizures in TUSZ data. See our paper for more details.

On terminal, run the following:

conda env create -f eeg_gnn.yml
conda activate eeg_gnn

The preprocessing step resamples all EEG signals to 200Hz, and saves the resampled signals in 19 EEG channels as h5 files.

On terminal, run the following:

python ./data/resample_signals.py --raw_edf_dir <tusz-data-dir> --save_dir <resampled-dir>

where <tusz-data-dir> is the directory where the downloaded TUSZ v1.5.2 data are located, and <resampled-dir> is the directory where the resampled signals will be saved.

Note that the remaining preprocessing step in our paper --- Fourier transform on short sliding windows, is handled by dataloaders. You can (optionally) perform this preprocessing step prior to model training to accelerate the training.

Preprocessing for seizure detection and self-supervised pre-training:

python ./data/preprocess_detection.py --resampled_dir <resampled-dir> --raw_data_dir <tusz-data-dir> --output_dir <preproc-dir> --clip_len <clip-len> --time_step_size 1 --is_fft

where <clip-len> is 60 or 12.

Preprocessing for seizure classification:

python ./data/preprocess_classification.py --resampled_dir <resampled-dir> --raw_data_dir <tusz-data-dir> --output_dir <preproc-dir> --clip_len <clip-len> --time_step_size 1 --is_fft

For distance-based EEG graph, we provide the pre-computed adjacency matrix in ./data/electrode_graph. For correlation-based EEG graphs, adjacency matrices can be obtained from the respective dataloaders in ./data.

We also provide helper functions in ./graph_viz/graph_viz_utils.py to visualize the EEG graphs. See notebook ./graph_viz/eeg_graph_visualization.ipynb for examples of graph visualization.

To train seizure detection from scratch using distance-based EEG graph, run:

python train.py --input_dir <resampled-dir> --raw_data_dir <tusz-data-dir> --save_dir <save-dir> --graph_type combined --max_seq_len <clip-len> --do_train --num_epochs 100 --task detection --metric_name auroc --use_fft --lr_init 1e-4 --num_rnn_layers 2 --rnn_units 64 --max_diffusion_step 2 --num_classes 1 --data_augment

where <clip-len> is 60 or 12.

To use correlation-based EEG graph, specify --graph_type individual.

To use preprocessed Fourier transformed inputs from the above optional preprocessing step, specify --preproc_dir <preproc-dir>.

To train seizure type classification from scratch using distance-based EEG graph, run:

python train.py --input_dir <resampled-dir> --raw_data_dir <tusz-data-dir> --save_dir <save-dir> --graph_type combined --max_seq_len <clip-len> --do_train --num_epochs 60 --task classification --metric_name F1 --use_fft --lr_init 3e-4 --num_rnn_layers 2 --rnn_units 64 --max_diffusion_step 2 --num_classes 4 --data_augment --dropout 0.5

Similarly, <clip-len> is 60 or 12. To use correlation-based EEG graph, specify --graph_type individual. To use preprocessed Fourier transformed inputs from the above optional preprocessing step, specify --preproc_dir <preproc-dir>.

To train self-supervised next time period prediction using distance-based EEG graph, run:

python train_ssl.py --input_dir <resampled-dir> --raw_data_dir <tusz-data-dir> --save_dir <save-dir> --graph_type combined --max_seq_len <clip-len> --output_seq_len 12 --do_train --num_epochs 350 --task 'SS pre-training' --metric_name loss --use_fft --lr_init 5e-4 --num_rnn_layers 3 --rnn_units 64 --max_diffusion_step 2 --data_augment

Similarly, <clip-len> is 60 or 12. To use correlation-based EEG graph, specify --graph_type individual. To use preprocessed Fourier transformed inputs from the above optional preprocessing step, specify --preproc_dir <preproc-dir>.

To fine-tune seizure detection/seizure type classification models from self-supervised pre-training, add the following additional arguments:

--fine_tune --load_model_path <pretrained-model-checkpoint>

In addition, we provide pretrained model checkpoints in the folder pretrained.

If you use this codebase, or otherwise find our work valuable, please cite:

@article{tang2021automated,
      title={Automated Seizure Detection and Seizure Type Classification From Electroencephalography With a Graph Neural Network and Self-Supervised Pre-Training}, 
      author={Siyi Tang and Jared A. Dunnmon and Khaled Saab and Xuan Zhang and Qianying Huang and Florian Dubost and Daniel L. Rubin and Christopher Lee-Messer},
      year={2021},
      journal={arXiv preprint arXiv:2104.08336}
}