Genuinely Blind Identification of Sleep Spindles through Trispectral Modulation Analysis

Abstract

Sleep spindles, transient 11-16 Hz oscillatory bursts, are a defining electrographic feature of non-rem (NREM) sleep and a key biomarker of sleep physiology. A need for efficient and reliable identification of spindles motivates a large literature on automated detection algorithms. In this literature, annotation by trained sleep specialists remains the gold standard against which automated methods are trained, tuned and evaluated. However, inter-scorer agreement among experts is modest, which leaves a significant role for subjective judgment in the definition of a spindle. Finding objective, scorer-independent, criteria for identifying spindles remains an unresolved challenge. We report here a robust, highly specific, and previously unrecognized signature of spindle activity in the fourth-order spectrum (trispectrum), from which we identify the presence of spindles, characterize their waveforms, and obtain an optimal detection filter through a decomposition of the trispectrum (HOSD). Although it is a strictly blind, data-driven method, HOSD-based spindle identification and detection agrees well with expert annotation (median AUROC ~0.9), yet identifies many more events at the native threshold than both human scorers and comparison detectors. Many of these additional detections are confirmed as meeting AASM spindle criteria by four blinded specialists, demonstrating that spindle-like oscillatory bursting is prevalent below conventional human and automated detection thresholds. We observe that N2 sleep is distinguished principally by high-amplitude bursts, while low-amplitude bursting persists throughout NREM sleep, being globally suppressed only in REM sleep. We also describe robust identification of recording-specific spindle waveform properties such as frequency deceleration.