This project, conducted by Riccardo Lunelli, Riccardo Parola, and Gabriele Santini, focuses on classifying patient pathologies from EEG signals using a convolutional neural network (CNN), specifically EEGNet. The primary objective is to leverage deep learning techniques to analyze EEG patterns associated with different neurological conditions.
The dataset, derived from fifteen epilepsy patients, includes EEG signals recorded during a verbal working memory task. The data encompass various signal parameters and are preprocessed for consistency and clarity, including frequency standardization and artifact removal.
- Preprocessing: Applied filtering and artifact removal to create a homogenized dataset.
- Model Implementation: Utilized EEGNet, a pretrained CNN model, with modifications to improve performance.
- Classification Tasks: Focused on both multiclass (specific pathologies) and binary (presence of hippocampal sclerosis) classifications.
- Model Training: Employed PyTorch for training, using F1 score and accuracy as primary metrics.
- Transitioning to ReLU activation significantly enhanced model performance.
- The limited dataset size emphasizes the need for more extensive data for refined accuracy.
- Future work includes partitioning data for disease-focused learning and enhancing model generalizability.
The runnable code is in the file EEGNet_improved.ipynb.