Search arXiv⌕ Search

arXiv · 2611.09576

Expert Spindles in the Wild: Enriched Supervision for a Dense Detector

Abstract

Sleep spindles are a hallmark of healthy NREM sleep whose frequency, amplitude, density, duration and topography vary with brain maturation, age and disease. Their manual annotation is impractical and variable between scorers, yet the public references used to train automatic detectors cover these characteristics only partially: stage N2 only, a single central derivation, or short pre-selected segments. Measuring duration and density requires dense detectors that learn event boundaries from annotated onsets and offsets, and therefore inherit the protocol under which those boundaries were scored. We introduce Expert Spindles in the Wild (ESW), a stage-agnostic expert spindle annotation set covering 19 complete MASS-SS2 recordings ($\approx$150 h of continuous EEG, 29,819 events). A sleep neurophysiologist annotated the recordings using the full EEG montage, blinded to sleep-stage labels and prior spindle annotations. On the 15 nights shared with historical scorers, ESW recovers 92.7% of E1 and 73.7% of E2 events, while recovering 81.6% of the MODA expert-consensus events on the common excerpts. Despite stage-blind annotation, 99.0% of ESW events occur in N2/N3. Used to train a fixed dense detector on a frozen SleepFM encoder, ESW reaches an event-level $F_1$ of 0.680 on 166 unseen MODA subjects, compared with 0.635 for E2 ($ΔF_1$ = +0.045, paired-bootstrap 95% CI [0.024, 0.065]). ESW is thus useful supervision for cross-dataset detection, and closer annotation-level agreement with the MODA consensus did not translate into better transfer performance in this experiment. The annotations are released under the MASS data-use terms (https://doi.org/10.5281/zenodo.23894975).

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Olivier Pallanca, Alexis Bollengier, Abel Díaz Berenguer, Hichem Sahli, André Anjos. 2026-11-06. Expert Spindles in the Wild: Enriched Supervision for a Dense Detector. https://arxiv.org/abs/2611.09576

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

Robust Model Order Selection via Dithered Differential Step-Down Thresholding

We propose a dithered differential step-down thresholding (DSDT) algorithm for robust model order selection in stationary noise. Evaluating only exceedance counts over a decreasing threshold grid, the proposed algorithm stops when detecting noise extrema value clustering. We prove that deterministic threshold grids are inconsistent under lowered floor detection thresholds due to threshold grid misalignment error. We bound this error using randomized grid dithering. For arbitrary deterministic shifts, any joint noise distribution, and data-dependent thresholds, we obtain non-asymptotic bounds on the estimation risk and family-wise error rate, establishing DSDT as a universal false alarm suppressor for any single-threshold algorithm. When the scale sequence tends to zero or extremes cluster along the search grid, DSDT achieves asymptotic exact order recovery of separated deterministic components exceeding the noise maximum plus the grid step, even below the minimax threshold or without a detection threshold. Using multiple independent dithers reduces error bounds. We provide convergence and consistency analysis for an arbitrary number of dithers. This analysis establishes asymptotic bounds under extreme value laws and non-asymptotic bounds under nondecreasing hazard rates. The results extend to unseparated deterministic shifts, random shifts, unknown noise scales, and multi-dimensional search grids. We show that DSDT is robust to data quantization.

eess.SP↗

Learning Latent Fields in Multiobject Tracking

Multiobject tracking using measurements from arrays of wireless sensors is a key task in applications ranging from autonomous navigation to public safety. Many existing methods compress raw complex-valued signals into detections or intensity maps during preprocessing, thereby discarding potentially relevant information about the objects and the propagation environment. We propose a Bayesian framework for multiobject tracking that relates object states to raw complex-valued signals via a physics-based superpositional model. The model combines a physics-based array manifold, such as radar steering vectors or the replica vectors of matched-field processing, with a latent field. This latent field is a family of context-dependent signal-power distributions that captures the effects of the objects and the propagation environment and is shared across objects and time. It is parameterized by a conditional normalizing flow, which is trained during runtime without labeled data by maximizing an evidence lower bound (ELBO) derived from the Bayesian model. Learning and inference are thus coupled in a sequential framework. We evaluate the method in simulation and in real-world radio and acoustic experiments, with spatial scales ranging from meters to kilometers, and show that it outperforms detection-based methods.

eess.SP↗

Energy-Aware Vertically Layered Pinching-Antenna System Design with Probabilistic LoS Availability

This paper investigates a vertically layered pinching antenna (PA) system, in which dielectric waveguides deployed at different heights provide additional vertical spatial degrees of freedom for line-of-sight (LoS)-favorable PA placement. We consider a practical energy model that captures in-waveguide attenuation, base station-to-feed-point optical-fiber loss, and PA repositioning energy, along with a statistical LoS-availability model that depends on elevation and distance. We then jointly optimize the transmit beamforming, PA positions, time allocation, and transmission rates to maximize the expected successfully delivered frame-effective sum rate under per-user outage and quality-of-service constraints. To avoid enumerating the exponentially many LoS states, we derive a certified conditional-moment outage envelope for outage-guaranteed rate selection and a differentiable safe approximation of the chance constraints to facilitate tractable optimization. Building on these results, a block coordinate descent algorithm with a penalty-aided majorization-minimization procedure attains a high-quality suboptimal solution. Simulation results confirm the gains of vertical layering and reveal a fundamental energy-LoS tradeoff: LoS-favorable PA placement generally enhances both performance and reliability, although it incurs higher attenuation, path loss, and energy costs.

eess.SP↗