CNS*2026 Halifax

Introduction
Sleep spindles —oscillatory bursts of 12-16 Hz associated with non-REM (NREM) sleep— are a key marker of sleep-related neuroplasticity. The gold standard for sleep spindle detection is commonly taken to be visual inspection by a trained sleep specialist. No blind, purely data-driven, methods have yet been demonstrated that are capable of identifying spectral and temporal properties of spindles independently of human judgment. Because inter-rater agreement among experts is modest, the lack of an objective criterion for identifying spindles leaves many open questions about their true prevalence and nature. Here we describe strictly data-driven identification and detection of spindles using information in the trispectrum.

Methods
A subdomain of the trispectrum was used to identify the presence and characteristics of modulated carrier oscillations and their envelopes by way of an interpretable representation (modulogram) [Kovach et al., 2026]. A decomposition of the trispectrum (HOSD) [Kovach and Howard, 2019] allowed separation of modulated oscillations from each other and background noise, and their characterization by a recovered feature waveform, conveying average spectral and temporal characteristic of oscillatory bursts. These methods were applied towards identifying spindles in two open databases containing polysomnography samples and expert annotation of spindles: MASS [Lacourse et al., 2020] (N=100) and DREAMS [Devuyst et al., 2011] (N=8).

Results


1) Tricoherence between 11Hz and 16 Hz (FDR Q ≪ 0.05) provided a highly robust and specific correlate of sleep spindles in every sample.
2) Oscillatory bursts identified with HOSD agreed well with expert spindle annotation (median AUC > 0.85), albeit at a lower amplitude threshold resulting in many more detections.
3) A high proportion of these additional detections were validated as true positives by 4 blinded sleep specialists.
4) N2 is distinguished by high-amplitude (> 12 dB) spindles while low amplitude oscillatory bursts in the spindle band are prevalent in all NREM stages.
5) HOSD feature identification reveals descending frequency ( median −3.9 Hz/s, IQR 1.6) as a characteristic of spindle waveforms (signed rank test, P≪0.001).



Discussion


HOSD applied to the trispectral modulogram provides a reliable means to spindle identification and detection that is (1) independent of human judgment, (2) highly robust as gauged against available reference data sets (3) capable of revealing novel insights into properties of spindle waveforms and their association with sleep state.

References


Devuyst, S., et al. (2011). Automatic sleep spindles detection—overview and development of a standard proposal assessment method. In 2011 Annual international conference of the IEEE engineering in medicine and biology society, pp. 1713–1716. IEEE.
Kovach, C. K., et al. Interpreting the trispectrum as the cross-spectrum of the wigner–ville distribution. IEEE Signal Processing Letters 33, 221–225.
Kovach, C. K. and M. A. Howard (2019).  Decomposition of higher-order spectra for blind multiple-input deconvolution, pattern identification and separation. Signal Processing 165, 357 – 379.
Lacourse, K., et al. (2020). Massive online data annotation, crowdsourcing to generate high quality sleep spindle annotations from eeg data. Scientific data 7 (1), 190.

Acknowledgement


\n\nNIH (Grant Number: 3UH3NS113769, R01NS117753 and R01DC004290)\n\nDOD (Grant Number: W81XWH-19-1-0637)