Subject-Relative Micro-Motion and Sleep Dynamics for Near-Infrared Video Sleep Staging
Kunmin Jang, You Rim Choi, Hun Heo, Heonjun Lee, Suahn Bae, Dongik Park, Hyun-Woo Shin, Hyung-Sin Kim
cs.CV · cs.AI
Quick verdict
ViNUSS: NIR-video-only sleep staging hits 0.80 Acc/0.78 MF1 on 475 nights; convincing ablations, but no significance tests and a 4-class headline.
Summary
The paper proposes ViNUSS, a framework that performs sleep staging directly from near-infrared (NIR) video, using PSG only as the label source rather than reconstructing respiratory/cardiac proxies or distilling EEG representations. It combines three components: a Pixel-Wise Temporal Encoder (PTE) that applies short temporal convolutions per pixel before spatial aggregation to preserve weak localized micro-motion; an Intra-Subject Discriminability Loss (ISDL), a triplet objective sampled within one night to learn stage cues relative to each subject's baseline; and two-scale Sleep Dynamics Modeling built on Mamba, where ShortMamba summarizes within-epoch motion and bidirectional LongMamba models the full-night epoch sequence. On the Korea Video Sleep Study (475 overnight recordings, ~3,250 hours, one hospital, 74% male), ViNUSS reports 0.80 accuracy, 0.68 Cohen's kappa, and 0.78 macro-F1 for four-class staging (trained with five AASM classes), exceeding oracle-input SleepVST variants (0.73 macro-F1 with full contact-sensor oracle) and video backbones (MoViNet-A2 baseline 0.50 macro-F1). Ablations attribute the largest single gain to SDM (+0.17 macro-F1); ISDL alone adds only ~0.01. SVECAM saliency highlights thoraco-abdominal motion. Key limitations: no significance testing or seed variance, a single data split, cross-dataset comparisons inherited from prior papers, and code only promised upon acceptance.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- NIR video alone, without respiratory/cardiac proxy reconstruction or EEG supervision, supports PSG-defined sleep staging at 0.80 accuracy / 0.78 macro-F1 (4-class) on 475 KVSS nights.
- Pre-spatial micro-motion encoding (PTE) preserves weak localized temporal cues that generic video backbones dilute during spatial aggregation.
- Within-subject triplet contrast (ISDL) reduces subject-dominant representation bias by learning stage cues relative to each night's baseline.
- Two-scale Mamba dynamics (ShortMamba + bidirectional LongMamba) yields coherent full-night trajectories and outperforms RNN, Transformer, and TCN alternatives on identical epoch features (Table D.3).
- SVECAM saliency indicates the model exploits physiologically meaningful evidence, chiefly thoraco-abdominal periodic motion and arousal-related gross movements.
Strengths
Realistic large-scale clinical corpus
475 overnight recordings / ~3,250 h of NIR video captured under routine sleep-lab conditions — far-field 3.5 m viewpoint, bedding occlusion, face anonymization (Appendix A) — an order of magnitude larger than prior camera studies listed in Table 1 (20–50 recordings).
Evidence-driven design motivation
Figure 1 (pixel-wise FFT vs contact chest-belt spectra; mismatch epochs like REM 322nd/345th) and Figure 2 (subject-dominant t-SNE clustering; MoViNet attending to static bedding folds) concretely demonstrate the proxy-unreliability and subject-bias failure modes the method targets.
Unusually thorough ablation suite
Honest baseline framing
Internal consistency of reported numbers
Weaknesses
No uncertainty quantification
Four-class headline inflates results
Headline component contributes least
Weak external comparison on KVSS
Generalizability and clinical ceiling unestablished
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ViNUSS addresses video-only sleep staging: PSG supplies the labels, but there is no respiratory/cardiac proxy reconstruction and no EEG-teacher supervision — a deliberate contrast to proxy-first (SleepVST, Video-PSG) and cross-modal (IoT-V2E, EEG distillation) lines in §2.2. Central claims: (1) NIR video alone is independently informative under PSG-defined labels; (2) weak local micro-motion must be encoded before spatial aggregation (PTE); (3) stage cues are subject-relative (ISDL); (4) two-scale Mamba dynamics organize epoch evidence into full-night trajectories. Headline: 0.80 Acc / 0.68 κ / 0.78 MF1 on 475 KVSS nights (Table 1); the 5-class counterpart is 0.71 / 0.62 / 0.72 (Table D.1a).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ViNUSS and how accurate is it?
ViNUSS is a deep learning framework that assigns AASM sleep stages (Wake, Light, Deep, REM) directly from near-infrared video recorded in a dark sleep lab, without using any physiological signals as input. On 475 overnight recordings from the Korean KVSS dataset, it reports 0.80 accuracy, 0.68 Cohen's kappa, and 0.78 macro-F1 for four-class staging (the model is trained with five AASM classes and N1/N2 are merged for evaluation).
Does ViNUSS need breathing or heartbeat signals?
No. Unlike proxy-first methods such as SleepVST or Video-PSG, ViNUSS does not reconstruct respiratory or cardiac waveforms from video and does not use EEG-based supervision; PSG is used only to provide the training labels. The authors argue NIR video motion itself carries stage evidence, supported by saliency maps concentrating on thoraco-abdominal breathing motion.
What data was the paper evaluated on, and is it public?
The Korea Video Sleep Study (KVSS), distributed via AI Hub Korea: 475 overnight near-infrared recordings (~3,250 hours) synchronized with clinical PSG from one hospital, with faces blurred. Access requires an application. Subjects are 74% male with a wide range of OSA severity, split 335/70/70 by subject. The code is promised upon paper acceptance.
How does ViNUSS compare with SleepVST?
On the same KVSS data, ViNUSS reaches 0.78 macro-F1 versus 0.73 for a full-oracle SleepVST variant that uses real contact-sensor breathing and cardiac waveforms (and 0.37 for a partial oracle). However, the SleepVST comparison used frozen pretrained encoders with a Random Forest head, so it tests one architecture under one transfer protocol rather than proving video universally beats physiological proxies.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 4's confusion-matrix panels cannot be mapped unambiguously to their (a)/(b)/(c) captions in the extraction; values were reconciled with the running text and Table D.3 (LongMamba-only: Deep P/R 0.63/0.56; full SDM: 0.67/0.68), which the extraction ordering appears to swap.
- Plot-internal content that exists only as graphics (t-SNE geometry in Fig. 2b, per-subject dot distributions in Fig. D.1, hypnograms in Fig. D.2) is assessed via captions and running text; no figure pixel values beyond those quoted in text were recoverable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.