Robust epilepsy classification using efficient CNNs with temporal windowing and dataset-independent learning

Timely diagnosis of epilepsy is of paramount importance for the effective management of the disease, given its status as one of the most common neurological disorders. Manual interpretation of EEGs is labor-intensive and subjective, emphasizing the necessity for automated diagnostic systems. The aim of this study is to introduce a robust and lightweight framework of convolutional neural networks (CNN) for detecting epilepsy classes from EEG signals. Using multi-scale temporal windowing and Continuous Wavelet Transform (CWT) for feature extraction capable of detailed time-frequency features, followed by processing through a lightweight CNN model architecture using depth-wise separable convolutions. The generalization capability is evaluated via cross-dataset validation, where we train on the Turkish EEG dataset and use external datasets CHB-MIT and Bonn for testing. The Turkish dataset achieves a classification accuracy of 94.8% according to the proposed model, while retaining excellent performance on datasets that have not been seen before, validating its versatility against distribution shifts. Finally, real-time performance assessment on a Raspberry Pi 4 establishes the potential of the method in embedded & portable diagnostic pipelines, with average inference time well below 30 milliseconds. These findings are important as they demonstrate the practical potential to deploy an accurate, generalizable, and near-real-time EEG-based seizure detection system that is relevant to clinical and wearable applications.