Reproduction of Real-Time Seizure Detection using Electroencephalogram (EEG)

Primary Contributions

Rohit Rao 12/07/2025

• denoising.py - Implements 3 frequency-domain transforms (Bandpass, Notch, SWT) to denoise EEG input signals before inference using the models defined in this repo.
• process_TUH_dataset.py - Adapted the original preprocessing code with improved structure and updates for modern libraries. This script streamlines the creation of an EEG signals processor for any EEG dataset.

I made substantial changes to the existing preprocessing, training, and inference code, including major bug fixes, quality-of-life adjustments, and new feature add-ons.

---

Setup

Install the required dependencies:

pip install -r requirements.txt

Preprocessing:

python 1_preprocess.py --data_type train --cpu_num 12 --label_type tse_bi --save_directory out/ --samplerate 200
python 1_preprocess.py --data_type dev --cpu_num 12 --label_type tse_bi --save_directory out/ --samplerate 200

Training:

# Model 1: ResNet Short LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./2_train.py \
    --project-name resnet_short_lstm_raw \
    --model resnet_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --epoch 8 \
    --batch-size 32 \
    --seizure-wise-eval-for-binary True

# Model 2: ResNet Dilation LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./2_train.py \
    --project-name resnet_dilation_lstm_raw \
    --model resnet_dilation_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --epoch 8 \
    --batch-size 32 \
    --seizure-wise-eval-for-binary True

Testing:

# Model 1: ResNet Short LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./3_test.py \
    --project-name resnet_short_lstm_raw \
    --model resnet_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --seed 0 \
    --denoise bandpass \
    --ignore-model-speed \
    --cpu 1

# Model 2: ResNet Dilation LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./3_test.py \
    --project-name resnet_dilation_lstm_raw \
    --model resnet_dilation_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --seed 0 \
    --denoise bandpass \
    --ignore-model-speed \
    --cpu 1

Seizure-wise Test

# Model 1: ResNet Short LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./4_seiz_test.py \
    --project-name resnet_short_lstm_raw \
    --model resnet_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --seizure-wise-eval-for-binary True \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --best \
    --seed 0 \
    --denoise bandpass

# Model 2: ResNet Dilation LSTM Raw
CUDA_VISIBLE_DEVICES=0 python3 ./4_seiz_test.py \
    --project-name resnet_dilation_lstm_raw \
    --model resnet_dilation_lstm \
    --task-type binary \
    --optim adam \
    --window-size 4 \
    --window-shift 1 \
    --eeg-type bipolar \
    --enc-model raw \
    --binary-sampler-type 6types \
    --binary-target-groups 2 \
    --seizure-wise-eval-for-binary True \
    --best \
    --seed 0 \
    --denoise bandpass

*Note: "bandpass can be swapped with "notch" or "swt".

Original Repository README:

Real-Time Seizure Detection using Electroencephalogram (EEG)

This is the repository for "Real-Time Seizure Detection using EEG: A Comprehensive Comparison of Recent Approaches under a Realistic Setting". You can checkout pdf file of our paper in our github repo.

• If you have used our code or referred to our result in your research, please cite:

@inproceedings{lee2022real,
  title={Real-Time Seizure Detection using EEG: A Comprehensive Comparison of Recent Approaches under a Realistic Setting},
  author={Lee, Kwanhyung and Jeong, Hyewon and Kim, Seyun and Yang, Donghwa and Kang, Hoon-Chul and Choi, Edward},
  booktitle={Conference on Health, Inference, and Learning},
  pages={311--337},
  year={2022},
  organization={PMLR}
}

Real-Time EEG Detection

We downsample the EEG signal and extract features. The models detect whether ictal / non-ictal signal appears within the 4-second sliding window input. We present an example case with Raw EEG signal but other signal feature extractors can also be applied in the pipeline.

Requirements

• PyTorch 1.8.1+cu102

To install all the requirements of this repository in your environment, run:

pip install -r requirements.txt

Preprocessing

To construct dataset with TUH EEG dataset, you can download EEG Seizure Corpus in the website and save it to $PATH_TO_EEG and run:

python preproces.py --data_type train --cpu_num *available cpu numbers* --label_type  *tse or tse_bi* --save_directory *path to save preprocessed files* --samplerate *sample rate that you want to re-sample all files*

Before running the code, please change $PATH_TO_EEG/ to your own path to EEG raw dataset.

Model Training

Check our builder/models/detection_models or builder/models/multiclassification repository to see available models for each task. To train the model in default setting, run a command in a format as shown below :

CUDA_VISIBLE_DEVICES=*device number* python ./2_train.py --project-name *folder name to store trained model* --model *name of model to run* --task-type *task*

For sincnet settin, add --sincnet-bandnum 7

Example run for binary seizure detection:

CUDA_VISIBLE_DEVICES=7 python3 ./2_train.py --project-name alexnet_v4_raw --model alexnet_v4 --task-type binary --optim adam --window-size 4 --window-shift 1 --eeg-type bipolar --enc-model raw --binary-sampler-type 6types --binary-target-groups 2 --epoch 8 --batch-size 32 --seizure-wise-eval-for-binary True

CUDA_VISIBLE_DEVICES=7 python3 ./2_train.py --project-name cnn2d_lstm_raw --model cnn2d_lstm_v8 --task-type binary --optim adam --window-size 4 --window-shift 1 --eeg-type bipolar --enc-model raw --binary-sampler-type 6types --binary-target-groups 2 --epoch 8 --batch-size 32 --seizure-wise-eval-for-binary True

Example run for SincNet signal feature extraction :

CUDA_VISIBLE_DEVICES=7 python3 ./2_train.py --project-name alexnet_v4_raw_sincnet --model alexnet_v4 --task-type binary --optim adam --window-size 4 --window-shift 1 --eeg-type bipolar --enc-model sincnet --sincnet-bandnum 7 --binary-sampler-type 6types --binary-target-groups 2 --epoch 8 --batch-size 32 --seizure-wise-eval-for-binary True

Other arguments you can add :

1. enc-model : preprocessing method to extract features from raw EEG data (options: raw, sincnet, LFCC, stft2, psd2, downsampled) psd2 is for Frequency bands described in our paper stft2 is for short-time fourier transform
2. seizure-wise-eval-for-binary : perform seizure-wise evaluation for binary task at the end of training if True
3. ignore-model-summary : does not print model summary and size information if True model summary is measured with torchinfo Please refer to /control/config.py for other arguments and brief explanations.

Model Evaluation

We provide multiple evaluation methods to measure model performance in different perspectives. This command will measure the model's inference time in seconds for one window.

python ./3_test.py --project-name *folder where model is stored* --model *name of model to test* --task-type *task*
python ./4_seiz_test.py --project-name *folder where model is stored* --model *name of model to test* --task-type *task*

Test and measure model speed

To evaluate the model and measure model speed per window using cpu, run the following command :

CUDA_VISIBLE_DEVICES="" python ./3_test.py --project-name *folder where model is stored* --model *name of model to test* --cpu 1 --batch-size 1

For sincnet setting, add --sincnet-bandnum 7 4_seiz_test.py is for evaluation metrics of OVLP, TAES, average latency, and MARGIN

Other arguments you can add :

1. ignore-model-speed : does not calculate model's inference time per sliding window if True