TASA: Transition-Aware Sparse Attention for Parameter-Efficient Sleep Stage Classification

Automated sleep stage classification is critical for scalingsleep medicine, yet many state-of-the-art models remaincomputationally heavy for wearable deployment. Fur-thermore, the N1 stage remains a persistent challengedue to its transitional nature. In this work, we proposeTransition-Aware Sparse Attention (TASA), a parameter-efficient architecture (≈ 214k parameters) that utilizesSparse Attention and a novel transition-aware multitasklearning head. Utilizing a three-channel configuration(EEG Fpz-Cz, EEG Pz-Oz, and horizontal EOG), ourmodel achieves a Cohen’s Kappa of 0.7427 and an over-all accuracy of 80.79% on the Sleep-EDF dataset. Criti-cally, TASA achieves a significant improvement in N1 re-call (0.6482) by treating temporal boundaries as a struc-tural regularizer. Our results demonstrate that localizedattention with linear complexity O(L · W ) offers a viablepath for high-performance sleep monitoring on resource-constrained wearable devices.