Electroencephalography (EEG) for epilepsy inevitably contains artifacts that may obscure epileptiform discharges or, in some cases, provide valuable diagnostic clues. Accurate classification of different artifact types is therefore clinically important. This paper proposes a hybrid artifact classification method that integrates deep learning and handcrafted features. A CNN-CBAM-LSTM network is employed to capture temporal features, while time-domain, frequency-domain, and entropy-based handcrafted features are extracted to complement the representation. The combined feature set is refined using the Relief-F algorithm to reduce redundancy and enhance discriminability, followed by a random forest classifier for final recognition. Experiments were conducted on long-term clinical EEG recordings, including six artifact types: eye movement, muscle, chewing, electrode, blinking, and rhythmic chewing. Results show that the proposed method achieves 97.8% classification accuracy, surpassing existing approaches. By leveraging both the automatic feature learning capability of deep models and the interpretability of handcrafted features, the method demonstrates high accuracy and strong generalization potential for artifact classification in epileptic EEG analysis.

Machine learning,Deep learning,Feature extraction,Feature selection,Classification