RASR-Net: Lightweight Cross-Subject EEG Emotion Recognition for Hearing-Impaired Individuals

Abstract

Cross-subject electroencephalography (EEG) emotion recognition for hearing-impaired individuals is limited by inter-subject distribution shifts, the acquisition burden of high-density systems, and overfitting on small cohorts. This study proposes RASR-Net, a Riemannian-anchored spectral-region lightweight residual network that combines unlabeled within-subject Euclidean alignment, robust five-band statistics, frontal-parietal attention and gated fusion, and a constrained neural correction to a Riemannian logistic anchor. The model was evaluated on 30 hearing-impaired participants using strict leave-one-subject-out validation, with model selection confined to training subjects and three random seeds (90 outer folds). RASR-Net achieved 58.563% accuracy and 54.859% macro-F1 with 7,227 trainable neural parameters. It outperformed same-channel SGD logistic regression by 6.07 percentage points in accuracy and EEGNet by 10.35 points. Using 23 frontal-parietal channels retained 98.94% of the accuracy of a parameter-matched 62-channel counterpart while reducing electrodes by 62.90%, the FLOPs lower bound by 58.09%, and single-thread CPU latency by 47.02%. These results indicate that geometry-anchored residual learning can improve robust cross-subject decoding while preserving a practical lightweight acquisition configuration.