Prediction of epileptic seizures in advance using EEG signals

Epilepsy is a chronic neurological disorder that arises from abnormal electrical activity in the brain, leading to sudden seizures that may cause severe consequences such as loss of consciousness, uncontrolled muscle movements, and communication difficulties. The unpredictable and abrupt nature of these seizures poses a serious threat to patients' quality of life and safety, potentially resulting in accidents, injuries, and even death. Therefore, accurately predicting epileptic seizures in advance is of critical importance to enable timely interventions and mitigate potential adverse outcomes. Recently, deep learning-based approaches have shown remarkable performance in the detection, classification, and prediction of epileptic seizures, outperforming traditional methods in complex biomedical signal analyses. These methods are particularly effective in capturing latent patterns and deep features in EEG signals. In recent years, the early and accurate prediction of epileptic seizures has become a key research area for enhancing patient safety and improving quality of life. In this context, deep learning methods have been widely employed to analyze EEG signals effectively. This thesis proposes innovative hybrid and fusion-based approaches to predict epileptic seizures several minutes before onset, utilizing the temporal representation of EEG signals. The proposed methods focus on capturing complex and hidden patterns within EEG signals more effectively for accurate seizure prediction. The proposed models—two of which serve as baseline architectures—are as follows: (a) Temporal Convolutional Network captures temporal dynamics of EEG signals using time-series analysis; (b) The Tiny Transformer processes EEG spectrograms through a low-complexity Transformer structure, offering a compact yet effective representation; (c) The Hierarchical Temporal Tiny Transformer hierarchically learns multi-scale temporal dependencies in EEG data using stacked temporal convolution and Transformer modules; and (d) The Parallel Temporal Tiny ViT Transformer processes EEG signals in two parallel branches—spectral and temporal—using convolutional and Transformer-based paths, which are later fused via a dedicated fusion mechanism. The hierarchical architecture processes temporal features in sequential modules: the TCN captures short-term dynamics, GRU structures the sequential data, temporal attention highlights informative time steps, and Tiny Transformer captures long-range dependencies through self-attention. In contrast, the parallel model processes raw EEG and spectrogram representations simultaneously using TCN and Vision Transformer (ViT) branches, respectively, and fuses complementary temporal and spectral features using a cross-attention-based strategy. While all proposed models demonstrated strong performance in experimental evaluations, the HiTTFormer architecture outperformed the others, achieving 93,32% sensitivity, 94,29 % precision, 96,19 % accuracy, and an AUC of 0.99. These results indicate that integrating different representations of EEG signals is a powerful strategy for seizure prediction. Moreover, the architectural innovations proposed in this study are expected to contribute uniquely to the existing literature. Overall, the findings confirm that the proposed approaches are effective in predicting epileptic seizures earlier and more accurately, offering valuable insights for the development of next-generation predictive systems.