The Generalized Sleep Spindles Detector: A Generative Model Approach on Single-Channel EEGs

Abstract

We propose a data-driven, unsupervised learning framework for one of the hallmarks of stage 2 sleep in the electroencephalogram (EEG)—sleep spindles. Neurophysiological principles and clustering of time series subsequences constitute the underpinnings of methods fully based on a generative latent variable model for single-channel EEG. Learning on the model results in representations that characterize families of sleep spindles. The discriminative embedding transform separates potential micro-events from ongoing background activity. Then, a hierarchical clustering framework exploits Minimum Description Length (MDL) encoding principles to effectively partition the time series into patterns belonging to clusters of different dimensions. The proposed algorithm has only one main hyperparameter due to online model selection and the flexibility provided by cross-correlation operators. Methods are validated on the DREAMS Sleep Spindles database with results that echo previous approaches and clinical findings. Moreover, the learned representations provide a rich parameter space for further applications such as sparse encoding, inference, detection, diagnosis, and modeling.