SleepEffFormer: Efficient CNN-Transformer with Transition-Aware Smoothing for Single-Channel EEG Sleep Stage Classification

Abstract

Automated sleep stage classification based on single-channel EEG is a promising pathway to large-scale sleep monitoring outside the polysomnographic lab. This paper proposes SleepEffFormer TAS, an efficient and interpretable system that consists of a four-stride-block 1D CNN feature extractor, a two-layer pre-normalization Transformer encoder, and a non-parametric Transition-Aware Smoothing (TAS) layer that suppresses physiologically unrealistic transitions between predicted stages. When tested on the Sleep-EDF Expanded dataset using 78 all-night EEG recordings from the Fpz-Cz channel with a subject-wise held-out split, the proposed model achieves 83.9% accuracy, a 78.9% macro F1-score, and a Cohen's kappa of 0.765 with approximately 367K learnable parameters. This performance is comparable to AttnSleep while using 3-5 times fewer parameters. Ablation studies show that the Transformer encoder improves macro F1 by 6.6 percentage points over a CNN-based system, while the TAS layer provides an additional 1.8 percentage point improvement without introducing any trainable parameters. Weighted loss is also critical for improving classification of the minority N1 sleep stage. Attention maps generated by the proposed model reveal physiologically sensible sleep-stage-related EEG features.