Automatic Segmentation of Sleep Spindles: A Variational Switching State-Space Approach

Mingjian He, Proloy Das, Gladia Hotan, Patrick L. Purdon

Abstract: Sleep spindles are transient 12–16 Hz oscillations during non-rapid eye movement sleep. Automatic detection of these discrete spindle events in electroen-cephalography has been a long-standing technical challenge, motivated by recent studies showing the integral role of spindles in normal sleep physiology, memory consolidation, and neurodegenerative diseases such as Alzheimer's disease. In this paper, we present an unsupervised spindle segmentation method based on switching state-space models. Unlike previous spindle detectors that apply band-pass filtering followed by thresholding, our method employs a class of neural oscillators to construct a generative model for spindles. In its simplest form, this framework proposes two parallel regimes representing either slow waves or slow waves and spindles, combined with a switching observation process that determines the presence of the spindle component. We derive a generalized EM algorithm based on a variational Bayes approximation to perform efficient inference and learning with this model, extending previous work by introducing switching models with nested structures enabling joint parameter estimation across models, and by introducing a principled warm-start strategy for initialization. When applied to real sleep EEG recordings, our method reliably detects and extracts spindles even when the spindle waveforms are weak.