A jointly optimized hybrid framework for epileptic EEG artifact classification

Accurate classification of artifacts in long-term electroencephalography (EEG) recordings is essential for reliable automated epilepsy analysis, as artifacts may obscure epileptiform patterns and degrade downstream detection performance. However, existing methods still face challenges in complementary feature integration and coordinated optimization of feature selection and classifier parameters. To address the limited feature representation capability of handcrafted methods and the limited interpretability of deep learning approaches, this paper proposes a hybrid EEG artifact classification framework that integrates complementary feature representation with the joint optimization of feature selector and classifier parameters. Specifically, handcrafted temporal-spectral-entropy features are fused with deep features extracted by a lightweight ResNet-LSTM network. A novel Class Neighborhood Discrimination (CND) method is employed for discriminative feature selection, while an Adaptive Seagull Optimization Algorithm (ASOA) jointly optimizes the parameters of the CND feature selector and the Random Forest (RF) classifier within a unified optimization framework. Experimental results on the Temple University Artifact Corpus (TUAR) and clinical EEG recordings show that the proposed framework achieves an accuracy of 99.30% on the originally constructed dataset and maintains 98.15% accuracy on an additional recording-independent clinical test set. The proposed method also shows consistent performance across multiple evaluation metrics, including MCC, Kappa, and GDR, and outperforms representative feature selection methods, optimization strategies, and classifiers. These results demonstrate that the proposed framework effectively combines interpretability, discriminative feature representation, and joint optimization, providing an accurate and robust solution for epileptic EEG artifact classification.