Physiology-informed spectral-temporal learning with confidence-aware evaluation for single-channel EEG sleep staging

Abstract

Automated sleep staging from single-channel electroencephalography (EEG) may reduce acquisition burden. Limited spatial information and inter-subject variability, however, make prediction difficult. We evaluated a physiology-informed spectral-temporal framework for single-channel EEG sleep-stage classification. Each 30-s Fpz-Cz epoch was transformed into a log-power short-time Fourier transform (STFT) representation and modulated across five sleep-related bands. Depthwise-separable convolutions and efficient channel attention extracted local time-frequency patterns. A lightweight Transformer modeled a centered seven-epoch window for center-epoch prediction. Training combined class-balanced focal loss, neighboring-epoch supervision and a transition prior estimated from training labels. On a fixed subject-disjoint Sleep-EDF evaluation partition, the framework achieved 0.8469 accuracy, 0.7905 macro F1 and 0.7906 Cohen’s kappa. It yielded higher discrimination metric point estimates than the protocol-matched light CNN, whereas the TCN produced slightly higher accuracy and kappa. The framework yielded the lowest expected calibration error (ECE) point estimate among the three evaluated models (0.0320). Retraining-based ablations removed spectral modulation, transition-prior loss or neighboring-epoch supervision under the same protocol. Across seeds 42, 123 and 2024, accuracy and macro F1 averaged 0.8530 ± 0.0060 and 0.8006 ± 0.0098, respectively. Removing the 20% lowest-confidence epochs yielded 0.9167 accuracy among the remaining automated epochs at 80% coverage. Repeated random referral at the same coverage left accuracy near the full-test-set estimate, whereas confidence-guided referral retained higher accuracy among the automated epochs. Test-time spectral, context and noise stress analyses examined sensitivity of the trained model without retraining. These results support further evaluation of the framework as an offline or delayed pre-scoring aid. The small fixed test cohort and absence of external validation preclude broader reliability or clinical generalization claims. The findings should be interpreted as methodological evidence rather than clinical validation.