Task-Oriented Semantic Feature Transmission for Robust EEG Motor Imagery Decoding Under Additive White Gaussian Noise

Remote electroencephalography (EEG) systems require compact representations that remain useful when communication noise corrupts the transmitted message. We evaluated whether task-oriented residual refinement of filter bank common spatial pattern (FBCSP) features compressed by principal component analysis (PCA) improves four-class motor imagery (MI) decoding without increasing the transmitted dimension. The BNCI2014-001 dataset was assessed in nine subjects using bidirectional subject-specific cross-session evaluation. All methods transmitted K∈{16,32,64} power-normalized real values through additive white Gaussian noise (AWGN) at seven signal-to-noise ratios (SNRs) and a noise-free reference. Balanced accuracy was averaged over 20 paired noise realizations per noisy condition, and paired subject-level differences were evaluated with exact joint sign-flip max-|t| inference. At K=32, the proposed method achieved 41.10%, 49.46%, and 56.15% balanced accuracy at -10, -5, and 0 dB, compared with 38.63%, 46.28%, and 53.26% for conventional FBCSP-PCA transmission. Ten of the 24 semantic-versus-conventional comparisons were significant after family-wise max-|t| correction, including all nine comparisons at -10, -5, and 0 dB. Receiver-only controls closely reproduced conventional performance at all three message dimensions, whereas alternative loss weights, uniform-SNR training and selection, and removal of the 0 dB/noise-free reference-condition guard retained positive low-SNR gains. Overall, baseline-preserving task-oriented refinement improved MI decision robustness under severe AWGN without increasing the number of transmitted values.