Epilepsy detection using EEG data.
Epileptic activity can be recognized by a specialist as activity with much higher signal amplitude (see [Vila-Vidal et als]). Despite this observation, automatic classification is expected to be hindered by differences in distance between the epicenter and the measurement electrode. The models presented here were not given any a priori information to use the signal amplitude as feature. Despite this, the classification models perform quite well.
This repository contains three folders: "src" contains notebooks with classification models and modules for feature extraction, "res" contains pictures produced durng Exploratory Analysis and "dat" contains the raw data.
Reference:
M. Vila-Vidal, C. P. Enríquez, A. Principe, R. Rocamora, G. Deco, A. T. Campo "Low entropy map of brain oscillatory activity identifies spatially localized events: A new method for automated epilepsy focus prediction", Neuroimage 208 (2020)
About the dataset.
Each row contains the signal of one EEG electrode during 1 second. The dataset comes from: https://www.kaggle.com/datasets/harunshimanto/epileptic-seizure-recognition
It contains 11500 rows and 178 features. The response variable contains the category of the 178-dimensional input vector (5 classes, one of them being seizure activity).
About the problem.
The goal is to recognize an Epileptic seizure against the rest. A task of binary classification is treated in this repository.
Selected models.
The features are treated as a time series of 178 points. Three ways to extract features lead to different classifiers:
- baseline classifier using the raw data as features (notebook:
classification_raw_data.ipynb) - apply discrete wavelet transform, then use classical statistical quantites (percentiles, mean, entropy, number of zero crossigs etc); also estimate the spectrum (Welch estimator) and use the positions and values of first spectral peaks (notebook:
classification_wavelets_fourier.ipynb). - use the features extracted by
tsfel.time_series_features_extractor(notebook:classification_tsfel.ipynb).
A Random Forest classifier (50 estimators) is trained on each set of extracted features.
Results.
Each model's performance is measured by its f1-score for the seizure class on test set and the number of features used for classification. The performances of all 3 models seem quite correct, which is consistent with the observation that the normal and epileptic signals tend to present different amplitudes (notebook explore.ipynb).
| Feature extraction | f1-score | number of features |
|---|---|---|
| Raw data | 0.91 | 178 |
| Fourier, wavelets | 0.96 | 80 |
tsfel |
0.95 | 10 |
Possible improvement.
The solution can be improved by inspecting the nature of extracted features. This seems relevant for both Fourier and Wavelet feature extractor and for tsfel. In the latter case, this can be done by adding an output information to the wrapper provided by extract_tsfel_features.py .
Feedback and additional questions.
All questions about the source code should be adressed to its author Alexandre Aksenov:
- GitHub: Alexandre-aksenov
- Email: [email protected]