Riemannian Geometry-Based Cross-Subject Motor Imagery EEG Classification via Subject-Adaptive Prototype Selection


Demirbaş G., Yılmaz Ç. M.

2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA), Ankara, Türkiye, 21 - 23 Mayıs 2026, ss.1-8, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/ichora69329.2026.11537072
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.1-8
  • Karadeniz Teknik Üniversitesi Adresli: Evet

Özet

This paper proposes a Riemannian alignment framework for cross-subject motor imagery classification using the Leave-One-Subject-Out (LOSO) strategy, combining manifold-based geometry with multi-band spectral analysis and discriminability-guided band selection. The proposed framework applies a multi-band filter-bank decomposition of five frequency bands, selected through a systematic band reduction analysis, to capture frequency-specific discriminative patterns. Each trial is represented as a regularized spatial covariance matrix on the symmetric positive-definite (SPD) manifold, and subjectadaptive prototype (SAP) selection constructs a target-specific training pool by selecting the geometrically closest source subjects using the Riemannian geodesic distance. The selected source covariances are aligned to the target subject using ManifoldConsistent Alignment (MCA). A dual-expert ensemble combines a global wide-band geometry specialist with a multi-band spectral specialist, and their probabilistic outputs are fused via arithmetic mean to improve robustness across neurophysiological profiles. The framework was evaluated on the PhysioNet EEGMMIDB dataset, achieving a mean accuracy of 72.44% and a Cohen's Kappa of 0.450 across the first 10 subjects without using any labeled data from the target subject.