Advancing SDG 3 with Seizure Ia Tools: a Comparative Analysis of Different Deep Learning Models for Epileptic Seizure Detection

Objective: This study compares different deep learning models to automate epileptic seizure detection from EEG cerebral registrations. This work aims to synthesize architectures, datasets, preprocessing, evaluation protocols, performance, computational complexity, and translational considerations (reproducibility, leakage risk, patient‑level generalization, and clinical deployment) in order to enhance epilepsy diagnosis and ensuring more safety and autonomy for patients. This investigation directly addressing the daily life challenges imposed by epilepsy and aligning with Sustainable Development Goal 3 (Good Health and Well-being). Theoretical Framework: EEG-based seizure detection relies on neurophysiological principles where ictal, interictal, and pre-ictal states exhibit distinct temporal and spectral patterns. Signal processing theory underpins the use of multiresolution analysis (e.g., DWT) to capture nonstationary EEG dynamics. Deep learning models, grounded in universal approximation and hierarchical feature learning, leverage CNNs for spatial patterns, RNNs (LSTM/GRU) for temporal dependencies, and attention for feature weighting. Hybrid approaches combine these theories to enhance accuracy, generalization, and interpretability in seizure detection and prediction. Method: Different models are discussed in this study: first strategy is based on a 3D convolutional autoencoder (3D-DCAE) was employed for unsupervised feature learning combined with a Bi-LSTM classifier on minimally preprocessed EEG segments. The second one is relying on a Discrete Wavelet Transform (DWT) to extract time–frequency and nonlinear features, then fused them in a CNN-GRU model with attention for detection and prediction. The third strategy proposed a DWT-based feature extraction pipeline feeding a 1D CNN-LSTM architecture, optimized for multi-dataset generalization and computational efficiency. All studies used cross-validation on public datasets (CHB-MIT, Bonn, TUSZ) and reported high accuracy, with variations in feature engineering and model complexity. Results and Discussion: Hybrid approaches show promising results in epileptic seizure detection, especially when combined with effective feature extraction methods. However, challenges like dataset limitations, real-time applicability, and under-researched seizure types (e.g., absence seizures) remain a delicate task. Future work should focus on addressing these gaps to enhance diagnostic accuracy and usability in clinical and home settings. Research Implications: This work provides a practical framework for developing accurate epileptic seizure detection and prediction tools. The findings compare divers IA models for advanced neurological care technologies and protocols. Originality/Value: This review uniquely synthesizes three cutting-edge deep learning approaches for EEG-based seizure detection and prediction, highlighting methodological diversity and comparative performance. It integrates a theoretical framework with practical recommendations for reproducibility and clinical translation. Unlike prior surveys, it emphasizes cross-dataset generalization, attention mechanisms, and hybrid architecture in a single comparative analysis.