Subject-Calibrated Attention-State Classification during Digital Cognitive Tasks via MTFF: Band-Token Cross-Band Fusion on Single-Channel EEG

<title>Abstract</title> <p>In the digital era, learning and work increasingly rely on digital tasks, making objective attention monitoring a practical problem for human-computer interaction and educational technology. Portable consumer-grade single-channel EEG enables low-burden recording, but the signal is vulnerable to ocular, facial EMG, and motion artifacts, and task-derived labels may contain residual uncertainty. To address these constraints, we propose MTFF (Multi-band Temporal Feature Fusion), a classification framework for portable single-channel EEG. MTFF learns intra-band rhythmic patterns in parallel over five canonical bands (δ/θ/α/β/γ) within 0.5-40 Hz and compresses each band into a learnable band token. A lightweight Transformer then operates on six tokens (five band tokens plus one class token) to model cross-band complementarity while reducing the cost and overfitting risk of long time-point self-attention. Finally, a capacity-controlled linear decision head with sparse regularization constrains model complexity. Data from 50 healthy university participants performing a digital cognitive task were evaluated using a subject-wise stratified 60/40 trial-level split rather than a leave-one-subject-out protocol. MTFF achieved a subject-wise mean AUC of 0.9214 and accuracy of 0.8640 on held-out test trials; a pooled trial-level ROC analysis across all test trials yielded a test AUC of 0.8928.</p>