Generalized Tonic-Clonic Seizures Detection Using Deep Learning Techniques
Name: JUAN SEBASTIÁN CAMPOS MESA
Publication date: 31/03/2025
Examining board:
Name |
Role |
|---|---|
| DANIEL CRUZ CAVALIERI | Examinador Externo |
| EVANDRO OTTONI TEATINI SALLES | Presidente |
| KLAUS FABIAN CÔCO | Examinador Interno |
| PATRICK MARQUES CIARELLI | Coorientador |
Summary: Generalized tonic-clonic seizures (GTCS) pose serious health risks, including an increased likelihood of sudden unexpected death in epilepsy (SUDEP), postictal pulmonary edema (PPE), and traumatic injuries from falls or jerking movements. However, GTCS detection remains challenging due to the complex and variable nature of EEG signals. Traditional methods struggle with these variations, while deep learning remains as a gold standard for seizure detection, making it a powerful tool for GTCS detection. This study explores the detection of generalized tonic-clonic seizures (GTCS) using EEG data from the Temple University Seizure (TUSZ) dataset. To achieve this, three deep learning architectures are employed: Diffusion Convolutional Recurrent Neural Network (DCRNN), Long Short-Term Memory (LSTM), and Convolutional Densely Connected Gated Recurrent Neural Network (C-DRNN). The research evaluates the impact of loss functions and data augmentation on model performance. Dice Entropy (DE) loss proves to be the most effective for DCRNN, while Cross-Entropy loss optimally enhances LSTM and C-DRNN. Data augmentation plays a crucial role in improving generalization and robustness across all models. DCRNN achieves the highest improvement, with AUC increasing from 0.675 to 0.781. LSTM follows, rising from 0.598 to 0.778, while C-DRNN improves from 0.647 to 0.775. These results suggest that the proposed approach can contribute to future projects aimed at improving GTCS detection in patients. Additionally, this study highlights the potential for further research into alternative data augmentation techniques and advanced methods to enhance model performance in GTCS detection.
