Computationally Efficient and Generalizable Machine Learning Algorithms for Seizure Detection from EEG Signals

Automated seizure detection from scalp EEG is critical for epilepsy management, yet existing algorithms often struggle to balance computational efficiency, predictive performance, and generalizability across diverse patient populations. This study investigates the ROCKET framework and complementary state-of-the-art frameworks in time series classification and seizure detection tasks, including Detach-ROCKET, Detach Ensemble, STFT-based feature transform, catch22 features and EEGNet. Models were trained on the TUSZ dataset and rigorously evaluated on both TUSZ and the independent Siena dataset to assess inter-subject and cross-dataset generalizability. The Detach Ensemble model achieved strong performance with median event-wise F1 scores of 0.89 on TUSZ and 0.53 on cross-data evaluation, while maintaining low false positive rates and exceptional computational efficiency. These results demonstrate that ROCKET-based models are competitive in achieving strong predictive performance, computational efficiency, and generalizability promising for practical clinical deployment. The implementation is available at https://github.com/zheyun-shou/eeg_seizure.