PaperarXiv:2611.21040

Towards wearable-scale, transition-aware sleep staging for clinical populations

PreprintDated Nov 16, 2026Added Nov 17, 2026

  • SH

Authors: Suvadeep Maiti (UCL); Keya Sinha (IIIT Hyderabad / IISER Bhopal); Raju S. Bapi (IIIT Hyderabad)

arXiv 2611.21040v1 [cs.HC] (2026-11-16)

TL;DR

A 136M-parameter sleep foundation model (SleepGPT), fine-tuned with labels from only a quarter of patients, teaches a 31k-parameter causal CNN-GRU (122 KB, 75 KB at INT8) to stage sleep in real time from one EEG and/or one EOG channel. A new transition-aware distillation loss (Gcv-TAD) puts more weight on rare stage changes and less on those that vary between patients. In stroke, REM-behaviour-disorder and mixed sleep-disorder cohorts it adds +3.2 to +7.6 macro-F1 points over plain training on the same labels, with the biggest gains on N1.

Abstract

Automatic sleep staging models are predominantly developed on healthy sleepers and generalize poorly to clinical populations with fragmented sleep. Across three cohorts (patients with stroke, isolated REM sleep behaviour disorder, and other sleep disorders), we observe 2-3x more awakenings per hour, nearly twice as much N1, and less REM sleep than in healthy sleepers. We propose a label-efficient teacher-student framework for wearable-scale sleep staging. SleepGPT (136M parameters), adapted using labels from 25% of training patients, supervises a causal CNN-GRU (31k parameters; 122 KB) that stages EEG, EOG, or both in real time. We introduce Gcv-TAD, a new black-box transition-aware distillation loss. It emphasizes uncommon stage transitions while discounting those that vary across patients, using a prior estimated from labelled patients. In patient-wise 10-fold cross-validation, Gcv-TAD improves per-patient macro F1 over supervised training on the same labels by +3.2 to +7.6 points on average. Gains are largest on N1, and performance approaches full-label training.

Verbatim from arXiv (opens in a new tab) · arXiv metadata (abstracts are released under CC0)

Results on ChatPPG leaderboards

All leaderboards

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

Tasks & signals

Tasks

Sleep stagingPulse arrival time

Signals

Keywords

Sleep StagingKnowledge DistillationTeacher StudentSleepgptLabel EfficientTransition Aware LossClinical SleepStroke

Datasets used

How it works

How Method works: Setup → Method → TargetSetup: Signals (eeg, eog). Method: Method. Target: Sleep staging; Pulse arrival time.SetupSignals: eeg, eogSignalseeg, eogMethodMethodTargetSleep stagingSleep stagingPulse arrival timePulse arrival time

Cite this paper

Cite this paper with the fields recorded here. Check it against the publisher before submitting.

Maiti, S., Sinha, K., & Bapi, R. S. (2026). Towards wearable-scale, transition-aware sleep staging for clinical populations. arXiv 2611.21040v1 [cs.HC]. https://arxiv.org/abs/2611.21040
Maiti, Suvadeep, Keya Sinha, and Raju S. Bapi. 2026. "Towards wearable-scale, transition-aware sleep staging for clinical populations." arXiv 2611.21040v1 [cs.HC]. https://arxiv.org/abs/2611.21040.
Maiti, S., Sinha, K. and Bapi, R. S. (2026) 'Towards wearable-scale, transition-aware sleep staging for clinical populations', arXiv 2611.21040v1 [cs.HC]. Available at: https://arxiv.org/abs/2611.21040
Maiti, Suvadeep, et al. "Towards wearable-scale, transition-aware sleep staging for clinical populations." arXiv 2611.21040v1 [cs.HC], 2026. https://arxiv.org/abs/2611.21040.
S. Maiti, K. Sinha, and R. S. Bapi, "Towards wearable-scale, transition-aware sleep staging for clinical populations," arXiv 2611.21040v1 [cs.HC], 2026, [Online]. Available: https://arxiv.org/abs/2611.21040.
@misc{maiti2026towards,
  title         = {{Towards wearable-scale, transition-aware sleep staging for clinical populations}},
  author        = {Maiti, Suvadeep and Sinha, Keya and Bapi, Raju S.},
  year          = {2026},
  howpublished  = {arXiv 2611.21040v1 [cs.HC]},
  eprint        = {2611.21040},
  archivePrefix = {arXiv}
}

Source metadata: 2026

Record updated Nov 17, 2026

Added to the ChatPPG research hub