S5: Self-Supervised Learning Boosts Sleep Spindle Detection in Single-Channel EEG via Temporal Segmentation.

Abstract

Sleep spindles, characteristic waveforms of N2 sleep in EEG, are associated with various neural processes such as cognitive function. However, their identification relies on visual inspection by experts-a time-consuming, labor-intensive, and low inter-rater consistency process that impedes cutting edge spindle research. We introduce S5, an automatic method for sleep spindle detection employing a novel encoder-decoder architecture for time-series segmentation. A two-stage training paradigm, comprising task-agnostic pre-training followed by downstream fine tuning, ensures high-precision identification. S5 demonstrates robust and competitive performance on two public datasets. On the multi-expert annotated MODA dataset, our method outperforms the average human expert. We further conducted an exploratory analysis on a large-scale unlabeled dataset of over 7,000 recordings as a physiological sanity check. S5 offers a precise and efficient solution for automating spindle detection, thereby accelerating related research. An accompanying graphical toolbox makes our method accessible for simple and intuitive analysis.