HCI has increasingly explored biosignals as intuitive input. Among them, EOG provides reliable information on eye movement and is suitable for eye-written character interfaces. However, most prior studies focused on eye-opened data, which depend on vi...
HCI has increasingly explored biosignals as intuitive input. Among them, EOG provides reliable information on eye movement and is suitable for eye-written character interfaces. However, most prior studies focused on eye-opened data, which depend on visual feedback and are easily affected by blink artefacts. This study addresses these limitations by proposing a transformer-based EOG model that operates stably under eye-closed conditions. The model uses an Improved Tokenizer to capture local temporal patterns without positional encoding and a Temporal Feature Fusion Network (TFFN) to integrate global temporal dependencies. Window Warping augmentation is also applied to increase data diversity and improve generalisation. Experiments show accuracies of 75.60%(10-fold) and 62.50%(LOSO), outperforming CNN, BiLSTM, and CCT models by 3.4–7.1%p. Ablation results confirm that both the tokenizer and TFFN significantly enhance local and global feature learning. These findings indicate that eye-written characters can be recognised without visual feedback, supporting the development of future real-time, user-independent, non-visual HCI systems.