A Variational Mode Decomposition-based Framework with Parameter Optimization and Reference-guided Feature Analysis for Epileptic EEG Signals

İkizler N.

IEEE ACCESS, sa.1, ss.1-15, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1109/access.2026.3701597
  • Dergi Adı: IEEE ACCESS
  • Derginin Tarandığı İndeksler: Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.1-15
  • Karadeniz Teknik Üniversitesi Adresli: Evet

Özet

Epileptic EEG signals are inherently non-stationary and nonlinear, requiring robust signal decomposition and discriminative feature representations for reliable seizure analysis. Although Variational Mode Decomposition has attracted growing interest in epileptic EEG studies, most existing methods rely on fixed parameter settings and are limited to binary classification, which restricts robustness and generalizability. This study proposes a systematic, signal-driven VMD-based EEG classification framework that emphasizes parameter optimization, reference-guided feature design, and comprehensive multi-class evaluation. Using the Bonn EEG dataset, the framework is evaluated across all 26 possible class combinations, spanning binary to five-class classification scenarios. The effects of key VMD parameters (K and α) are systematically investigated through grid-based performance analysis to identify optimal parameter ranges. In addition to conventional time- and frequency-domain statistical features, novel reference-based features are introduced by quantifying spectral divergence and energy-related differences between decomposed signal components and class-specific reference signals, enabling the capture of relative and physiologically meaningful spectral characteristics. The resulting feature sets are evaluated using k-NN, SVM and RF classifiers under a strict 10-fold cross-validation protocol. Experimental results show that the proposed reference-based features consistently improve classification, accuracy and stability, particularly in higher-order multi-class problems. Statistical analyses confirm that the observed performance gains over baseline feature models are highly significant. Notably, the RF classifier achieves an average accuracy of 99.46% across all 26 class combinations, demonstrating strong generalization capability. Overall, the proposed framework offers an interpretable, computationally efficient and provides a promising framework for epileptic EEG analysis, with potential applicability to biomedical and wearable EEG systems.