Automated seizure detection based on scalp electroencephalography (EEG) can significantly accelerate the epilepsy diagnosis process. However, most existing deep learning-based epilepsy detection methods are deficient in mining the local features and global time series dependence of EEG signals, limiting the performance enhancement of the models in seizure detection. Our study proposes an epilepsy detection model, CMFViT, based on a Multi-Stream Feature Fusion (MSFF) strategy that fuses a Convolutional Neural Network (CNN) with a Vision Transformer (ViT). The model converts EEG signals into time-frequency domain images using the Tunable Q-factor Wavelet Transform (TQWT), and then utilizes the CNN module and the ViT module to capture local features and global time-series correlations, respectively. It fuses different feature representations through the MSFF strategy to enhance its discriminative ability, and finally completes the classification task through the average pooling layer and the fully connected layer. The effectiveness of the model was validated by experimental evaluations on the publicly available CHB-MIT dataset and the Kaggle 121 people epilepsy dataset. The model achieved 98.85% classification accuracy and other excellent metrics in single-subject experiments on the CHB-MIT dataset, and also demonstrated strong performance in cross-subject experiments on the Kaggle dataset. Ablation experiments demonstrate the complementary roles of the CNN and ViT modules, and their integration significantly improves detection accuracy and generalization. Comparisons with other methods highlight the advantages of the CMFViT model. The CMFViT model provides an efficient, accurate, and innovative solution for complex EEG signal analysis and seizure detection tasks for single and cross-subjects while laying the foundation for developing real-time, accurate seizure detection systems.

Background: Epileptic seizures are unpredictable, and while existing deep learning models achieve high accuracy, their deployment on wearable devices is constrained by high computational costs and latency. To address this, this work proposes the RGF-Model, a lightweight network that unifies seizure prediction and detection within a single causal framework. Methods: By integrating Feature-wise Linear Modulation (FiLM) with a Ring-Buffer Gated Recurrent Unit (Ring-GRU), the model achieves adaptive task-specific feature conditioning while strictly enforcing causal consistency for real-time inference. A multi-teacher knowledge distillation strategy is employed to transfer complementary knowledge from complex teacher ensembles to the lightweight student, significantly reducing complexity without sacrificing accuracy. Results: Evaluations on the CHB-MIT and Siena datasets demonstrate that the RGF-Model outperforms state-of-the-art teacher models in terms of efficiency while maintaining comparable accuracy. Specifically, on CHB-MIT, it achieves 99.54% Area Under the Curve (AUC) and 0.01 False Prediction Rate per hour (FPR/h) for prediction, and 98.78% Accuracy (Acc) for detection, with only 0.082 million parameters. Statistical significance was assessed using a random predictor baseline ($p$ < 0.05). Conclusions: The results indicate that the RGF-Model provides a highly efficient solution for real-time wearable epilepsy monitoring.