Estratégias computacionais para detecção de epilepsia em EEG : abordagem em pipeline versus end-to-end
Uloženo v:
| Hlavní autor: | |
|---|---|
| Datum vydání: | 2025 |
| Médium: | Bachelor thesis |
| Jazyk: | por |
| Zdroj: | Repositório Institucional da UPF |
| Download full: | https://repositorio.upf.br/handle/123456789/9862 |
Shrnutí: | Electroencephalography plays a pivotal role in diagnosing neurological disorders such as epilepsy. This study contrasts two computational strategies for seizure detection: a feature engineering pipeline utilizing the Binary Dragonfly Algorithm for feature selection followed by a Deep Neural Network, and an end-to-end architecture, the Hybrid Bidirectional Convolutional Network 5. Employing a paired experimental design with common random seeds on the Bonn dataset, the trade-off between interpretability and generalization was evaluated. Results indicate that while the pipeline approach achieved high sensitivity for ictal events, it exhibited instability in feature selection and lower specificity for the interictal state. Conversely, the hybrid architecture demonstrated superior statistical robustness and greater generalization capacity, significantly enhancing the detection of interictal patterns (95.11% recall versus 80.00%), a 15.11 percentage point improvement critical for clinical reliability. It is concluded that while pipelines offer granular explainability, identifying potential model biases, end-to-end models provide greater reliability for clinical screening and continuous monitoring. |
Buďte první, kdo okomentuje tento záznam!