Transfer learning for subject-independent motor imagery EEG classification using convolutional relational networks.

Motor imagery (MI) based electroencephalography (EEG) classification is central to brain-computer interface (BCI) research but practical deployment remains challenging due to poor generalization across subjects. Inter-individual variability in neural activity patterns significantly limits the development of subject-independent BCIs for healthcare and assistive technologies. To address this limitation, we present a transfer learning framework based on Convolutional Relational Networks (ConvoReleNet) designed to extract subject-invariant neural representations while minimizing the risk of catastrophic forgetting. The method integrates convolutional feature extraction, relational modeling, and lightweight recurrent processing, combined with pretraining on a diverse subject pool followed by conservative fine-tuning. Validation was conducted on two widely used benchmarks, BNCI IV-2a (four-class motor imagery) and BNCI IV-2b (binary motor imagery), to evaluate subject-independent classification performance. Results demonstrate clear improvements over training from scratch: accuracy on BNCI IV-2a increased from 72.22 (±20.49) to 79.44% (±11.09), while BNCI IV-2b improved from 75.10 (±17.17) to 83.85% (±10.30). The best-case performance reached 87.55% on BNCI IV-2a with Tanh activation and 83.85% on BNCI IV-2b with ELU activation, accompanied by reductions in inter-subject variance of 45.9 and 40.0%, respectively. These findings establish transfer learning as an effective strategy for subject-independent MI-EEG classification. By enhancing accuracy, reducing variability, and maintaining computational efficiency, the proposed framework strengthens the feasibility of robust and user-friendly BCIs for rehabilitation, clinical use, and assistive applications.