Skip to content

luizantoniona/eeg-epileptic-seizure-detection

BranchesTags

Repository files navigation

eeg

This project focuses on the detection of epileptic seizures in EEG data using various machine learning models. The models implemented aim to classify EEG signals using different data domains: temporal, frequency, and time-frequency.

Table of Contents

Releases and Publications

📌 NOTE: If you use this software in academic work, please cite the following papers:

🚀 Version 2.0 - Current main branch

@article{anghinoni2025use,
  title={The use of Multi-domain Electroencephalogram Representations in the building of Models based on Convolutional and Recurrent Neural Networks for Epilepsy Detection},
  author={Anghinoni, Luiz Antonio Nicolau and Denardin, Gustavo Weber and Gertrudes, Jadson Castro and Casanova, Dalcimar and Oliva, Jefferson Tales},
  journal={arXiv preprint arXiv:2504.17908},
  year={2025}
}

📌 Version 1.0

@article{anghinoni2024detecccao,
  title={Detec{\c{c}}{\~a}o de epilepsia em eletroencefalogramas utilizando redes neurais convolucionais reduzidas},
  author={Anghinoni, Luiz Antonio Nicolau and Teixeira, Marcelo and de Castro Barbosa, Marco Antonio and Casanova, Dalcimar and Oliva, Jefferson Tales},
  journal={Journal of Health Informatics},
  volume={16},
  number={Especial},
  year={2024}
}

Packages

The following Python libraries were used in this project:

  • wfdb – For download dataset files and info.
  • Tensorflow – For model training and evaluation.
  • MNE – For EEG signal processing.
  • Keras Tuner – For hyperparameter optimization.
  • PyQt – Used to develop the front-end interface for the application.
  • sqlite3 - Used to store info on database.

Graphical Application

The graphical application is developed using PyQt and provides an interface for users to interact with the EEG data, visualize results, and configure model parameters.

To run:

python3 main.py

Main Scripts

These scripts serve as entry points for users who are not using the graphical interface, and provide utility tools to configure the dataset, manage the database, train models, and generate useful visualizations:

  • Database Configuration — Initializes the database and processes metadata for the EEG recordings.
  • Model Training — Handles training of machine learning models using different data representations.

Datasets

CHB-MIT Scalp EEG Database

This project utilizes the CHB-MIT Scalp EEG Database from PhysioNet [1]. The dataset was collected at Boston Children’s Hospital and includes EEG recordings from 22 pediatric patients with intractable seizures.

Each subject includes 9 to 42 EEG recordings, with seizure occurrences annotated. A summary file is provided per patient with key metadata such as seizure start times, durations, and file identifiers.

Due to the volume and complexity of the files, a database solution is employed to streamline access and improve the management of seizure metadata across subjects.

Database

The database is implemented using sqlite3 and stores information about the EEG files, including metadata and seizure annotations. The database schema is designed to facilitate efficient querying and retrieval of data for preprocessing and model training.

summary_info:

Used to store information about the files in the dataset.

dataset_name record_name file_name start_time end_time nr_occurrence start_occurrence end_occurrence nr_channels ds_channels disease_type
CHBMIT ch01 chb01_01.edf 12:34:22 13:13:07 2 1862, 2000 1963, 2213 24 FP1-F7,F7-T7,... epilepsy

metrics_info:

Used to store information about the training sessions of models.

dataset_name model_name model_data_domain model_window accuracy precision sensitivity specificity true_positive_rate false_positive_rate f1_score
CHBMIT CNN Time 5 0.876 0.876 0.876 0.876 0.876 0.876 0.876

Preprocessing

Generate segmented data:

  • Search anomalous summaries: Search in the database for EEG files containing epileptic seizures.

  • For each file: (This steps are performed using multithreading for improved performance)

    • Read file: Read the EEG file.

    • Remove unused channels: Eliminate any channels that are not required for analysis.

    • Sliding window segmentation: Segment the data into smaller windows. These segments are centered around seizures to optimize time and memory usage.

    • Compute PSD (Power Spectral Density): If the requested data is in the frequency domain, compute the Power Spectral Density (PSD) of each segment.

    • Compute Spectrogram: If the requested data is in the time-frequency domain, compute the spectrogram of each segment.

  • Pre-processed data: Obtain the pre-processed data as the output of the entire process.

   flowchart LR
   A[Search anomalous summaries]
   A --> B
   B[Read file]
   subgraph B [For each file]
      C[Read file]
      D[Remove unused channels]
      C-->D
   end
   B-->E
   subgraph E [For each epileptic seizure]
      F[Sliding window segmentation]
   end
   E-->G
   G{Domain?}
   H[Compute PSD]
   I[Compute Spectrogram]
   J[Pre-processed data]
   G-->|Time| J
   G-->|Frequency Welch| H
   G-->|Frequency Multitaper| H
   G-->|Time-Frequency| I
   H-->J
   I-->J
Loading

Adjust data for training:

  • Pre-processed data: This represents the pre-processed data obtained from the previous stage of data pre-processing.

  • Normalization: The pre-processed data is normalized to ensure that all features have a similar scale, which helps improve the convergence speed of the training algorithm and prevents any particular feature from dominating the learning process.

  • Data split: The normalized data is split into two subsets: training data, and test data.

    • Train: This subset is splited in 80% used to train the machine learning model, and 20% to tune the hyperparameters.

      • 20%: Used for hyperparameter tuning

      • 64%: Used for training

      • 16%: Used for validations

    • Test: This subset is used to evaluate the performance of the trained model on unseen data.

  • Hyperparameter tuning: Optimization of model parameters to improve performance.

  • Training: This stage involves training the machine learning model using the training data.

  • Evaluation: After training, the model's performance is evaluated using the test datasets to acess its generalization ability and effectiveness.

Models and Training

Data Domains

The data will be processed and inserted into the models in three different domains:

  • Temporal:

    • Raw EEG.

  • Frequency:

    • Power Spectral Density - Welch method.
    • Power Spectral Density - Multitaper method.

  • Time-frequency

    • Spectrogram - Multitaper method.

Models:

  • CNN (Convolutional Neural Network)
  • RNN (Recurrent Neural Network)
  • CRNN (Convolutional Recurrent Neural Network)

Metrics:

A set of metrics was used to evaluate the performance of the models. For those metrics, we consider $TP$ for true positives, $TN$ for true negatives, $FN$ for false negatives, $FP$ for false positives, and $T$ as the total number of samples.

Used Metrics:

Metric Equation
Acuracy $Acu = \frac{TP + TN}{T}$
Precision $Pre = \frac{TP}{TP + FP}$
Sensitivity $Sen = \frac{TP}{FN + FP}$
Specificity $Spe = \frac{TN}{TN + FP}$
F1-Score $F1 = \frac{2 \times Acu \times Sen}{Acu + Sen}$

Evaluation Tables:

Evaluation of RNN

Metric Window Time Welch Multitaper Spectrogram
Accuracy 1s 78.49±2.43 84.78±0.46 87.86±0.49 83.47±1.48
2s 79.74±1.66 85.40±0.37 91.65±0.56 83.73±1.71
5s 77.29±0.84 92.96±0.49 95.99±0.19 80.91±1.10
10s 76.83±1.78 96.18±0.43 97.05±0.24 79.67±2.08
Precision 1s 79.26±3.30 85.64±2.02 89.45±1.54 86.61±2.90
2s 80.45±4.41 87.21±1.20 93.39±1.18 85.09±2.15
5s 75.06±4.41 94.09±1.22 96.59±0.58 82.08±3.30
10s 78.58±4.42 97.37±0.59 97.55±0.53 81.53±3.89
Sensitivity 1s 78.10±2.92 84.07±1.42 86.58±1.17 81.37±1.79
2s 79.11±2.30 84.20±1.15 90.28±1.22 82.88±2.10
5s 78.98±3.05 92.04±1.31 95.44±0.70 80.19±2.57
10s 75.50±3.16 94.96±0.77 96.47±0.52 78.17±3.89
Specificity 1s 79.05±2.78 85.61±1.50 89.27±1.27 85.97±2.49
2s 80.69±3.06 86.75±0.87 93.17±1.06 84.70±1.99
5s 76.17±2.34 93.99±1.11 96.56±0.54 81.95±2.20
10s 78.73±2.56 97.42±0.56 97.62±0.50 81.74±2.41
F1-Score 1s 78.29±2.59 84.41±0.83 87.21±0.74 82.40±1.50
2s 79.41±1.73 84.79±0.72 90.96±0.83 83.30±1.85
5s 78.10±1.72 92.50±0.85 95.71±0.42 80.53±1.70
10s 76.15±2.38 95.56±0.57 96.76±0.33 78.89±2.08

Evaluation of CNN

Metric Window Time Welch Multitaper Spectrogram
Accuracy 1s 82.01±0.81 84.70±0.32 88.51±0.24 84.04±0.57
2s 83.59±0.55 88.95±0.37 93.61±0.32 89.20±0.35
5s 86.71±0.47 94.07±0.26 96.82±0.24 86.34±0.64
10s 83.19±0.58 96.03±0.24 97.63±0.16 87.12±1.84
Precision 1s 84.39±1.84 86.93±0.87 89.76±1.57 87.71±1.92
2s 88.28±2.64 91.34±0.80 93.65±0.77 91.54±1.08
5s 91.11±1.22 95.24±0.77 97.91±0.36 93.02±1.92
10s 84.93±2.24 97.43±0.42 97.85±0.36 91.62±5.69
Sensitivity 1s 80.44±1.71 83.06±0.59 87.47±1.32 81.62±1.75
2s 80.77±1.42 87.19±0.74 93.58±0.60 87.46±1.04
5s 83.76±1.15 93.07±0.56 95.82±0.55 82.02±1.70
10s 81.61±1.92 94.62±0.34 97.33±0.43 83.84±3.30
Specificity 1s 83.84±1.26 86.49±0.70 89.66±1.24 86.96±1.48
2s 87.18±2.15 90.91±0.70 93.65±0.69 91.15±0.92
5s 90.29±1.06 95.14±0.72 97.87±0.35 92.12±1.79
10s 85.00±1.50 97.46±0.40 97.93±0.34 91.70±4.79
F1-Score 1s 81.21±1.20 83.87±0.40 87.98±0.74 82.81±1.15
2s 82.15±0.82 88.06±0.52 93.59±0.39 88.32±0.66
5s 85.21±0.78 93.56±0.33 96.32±0.38 84.12±1.15
10s 82.38±1.21 95.32±0.26 97.48±0.27 85.41±2.19

Evaluation of CRNN

Metric Window Time Welch Multitaper Spectrogram
Accuracy 1s 81.93±0.55 84.73±0.36 88.21±0.32 84.11±0.74
2s 85.63±0.57 87.01±0.48 92.38±0.37 88.28±0.69
5s 88.27±1.48 93.38±0.25 95.88±0.32 86.11±0.66
10s 90.08±1.64 95.54±0.49 96.87±0.32 86.74±1.56
Precision 1s 86.66±2.57 86.27±1.45 89.94±1.65 91.05±1.29
2s 88.19±3.49 88.48±0.93 93.15±0.84 91.65±1.90
5s 91.69±1.31 93.60±1.16 97.13±0.67 88.37±3.14
10s 93.43±3.41 96.07±0.86 97.98±0.47 88.21±4.31
Sensitivity 1s 79.09±1.98 83.57±1.26 86.84±1.12 79.84±1.41
2s 84.01±1.86 85.97±1.02 91.75±0.69 85.91±1.53
5s 85.88±2.27 93.21±0.97 94.76±0.45 84.82±1.49
10s 87.28±1.90 94.90±0.85 95.74±0.73 85.27±2.23
Specificity 1s 85.57±1.91 86.03±1.00 89.75±1.35 89.78±1.13
2s 87.73±2.69 88.14±0.74 93.05±0.75 91.11±1.66
5s 91.10±1.26 93.61±1.04 97.07±0.65 87.78±2.58
10s 93.37±3.14 96.19±0.79 98.02±0.44 88.59±3.44
F1-Score 1s 80.48±1.26 84.14±0.76 87.52±0.61 81.92±1.07
2s 84.80±0.86 86.49±0.71 92.06±0.47 87.08±1.02
5s 87.06±1.85 93.29±0.53 95.32±0.32 85.45±0.69
10s 88.65±1.55 95.22±0.61 96.30±0.50 85.98±1.56

References

[1] Guttag J. CHB-MIT Scalp EEG Database (version 1.0.0). PhysioNet. 2010. Available from: https://doi.org/10.13026/C2K01R

License

Copyright © 2025 Luiz Antonio Nicolau Anghinoni

About

EEG Epileptic Seizure Detection project that focuses on the detection of epileptic seizures in EEG recordings using machine learning models applied across multiple data domains: temporal, frequency, and time-frequency. It provides tools for data preprocessing, model training, evaluation, and an interactive graphical interface to explore and analyze

Topics

python deep-learning signal-processing cnn eeg statistical-analysis rnn eeg-signals signal-analysis signal-detection epilepsy eeg-analysis crnn eeg-classification eeg-signals-processing seizure-detection

Resources

Readme

License

CC-BY-4.0 license
Activity

Stars

6 stars

Watchers

2 watching

Forks

1 fork
Report repository

Releases 2

V2.0.0-Final Latest
May 9, 2025
+ 1 release

Contributors

Languages