Single-Channel EEG-Based Epileptic Seizure Prediction Using Common Spatial Pattern and Transfer Learning

Epilepsy is a neurological disorder characterized by recurrent seizures. Effective seizure prediction can significantly improve patient safety and quality of life. Current EEG-based methods for seizure detection often rely on multichannel setups, which increase patient discomfort and computational complexity. This paper proposes a patient-specific, single-channel seizure prediction model using the CHB-MIT dataset. The model employs Common Spatial Patterns (CSP) for channel selection and Continuous Wavelet Transform (CWT) scalograms for feature extraction. A ResNet50-based model is trained on these scalograms to distinguish between preictal and interictal states. The proposed method achieves an average accuracy of 85.1±3.2%, sensitivity of 84.0±3.9%, and specificity of 87.4±2.8% across 13 cases, demonstrating its potential for wearable and accessible seizure detection systems. Future work aims to validate the approach on larger datasets and explore patient-independent models.