SleepPairNet: Learning Multimodal Pediatric Sleep Representations with Random-Pair Cross-Attention

Saurav Raj Pandey; Harlin Lee

cs.LGOct 19, 2026

SleepPairNet: Learning Multimodal Pediatric Sleep Representations with Random-Pair Cross-Attention

Authors: Saurav Raj Pandey, Harlin Lee

Organizations: Department of Computer Science, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA · School of Data and Information Sciences, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA

Abstract

Analyzing physiological sleep signals is important for assessing sleep quality and detecting sleep-related disorders. Learning inter-modal relationships across physiological signals can provide richer representations for downstream sleep analysis. We introduce SleepPairNet, a multimodal self-supervised learning (SSL) framework that learns pairwise interactions between physiological signals through cross-attention. We propose a random modality-pair selection strategy that exposes the model to diverse signal pairs during pretraining, enabling flexibility across modality combinations at test time. Pretrained on the NCH dataset, SleepPairNet outperforms the evaluated SSL baselines on all respiratory tasks while remaining competitive on EEG-related tasks, and shows even stronger results when transferred to the external CHAT dataset.

Figures & tables

Figure 1: Overview of SleepPairNet’s architecture.

Model	Modalities	Performance (%)
		Sleep Scoring				Apnea (0.8%)				EEG Arousal (4.6%)				Oxygen Desat. (8.7%)				Hypopnea (1.9%)
		Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR
SleepFM	$c_{1},c_{2}$	66.8 (0.0)	60.1 (0.1)	88.5 (0.1)	64.7 (0.1)	98.2 (0.0)	14.8 (0.2)	80.4 (0.1)	7.3 (0.2)	92.8 (0.1)	41.7 (0.2)	87.4 (0.1)	36.0 (0.5)	84.8 (0.3)	36.2 (0.1)	77.8 (0.1)	31.9 (0.1)	96.3 (0.1)	17.5 (0.3)	77.8 (0.1)	10.7 (0.1)
PedSleepMAE	$c_{1},c_{2}$	60.5 (1.0)	46.2 (1.5)	84.1 (0.1)	53.5 (0.2)	97.8 (0.2)	11.4 (0.1)	74.1 (0.1)	4.4 (0.0)	87.4 (0.6)	24.4 (0.1)	75.3 (0.0)	13.9 (0.0)	81.3 (0.2)	27.0 (0.0)	69.7 (0.0)	21.1 (0.0)	94.0 (0.1)	14.9 (0.1)	73.0 (0.2)	7.2 (0.0)
SynthSleepNet	EEG, EOG	74.8 (0.3)	67.2 (0.2)	92.5 (0.0)	72.2 (0.0)	92.2 (5.5)	5.5 (0.4)	73.6 (1.1)	2.6 (0.3)	93.2 (0.1)	40.1 (0.2)	84.3 (0.3)	33.0 (0.2)	71.3 (0.9)	26.7 (0.1)	70.9 (0.1)	18.0 (0.1)	93.3 (0.3)	10.1 (0.4)	73.2 (0.4)	5.6 (0.2)
SynthSleepNet	EEG, EMG	73.2 (0.3)	65.9 (0.2)	91.9 (0.0)	71.2 (0.1)	95.2 (0.1)	8.3 (0.2)	77.0 (0.8)	3.7 (0.2)	93.0 (0.2)	41.6 (0.2)	86.8 (0.1)	35.4 (0.1)	71.9 (1.1)	27.1 (0.2)	71.5 (0.2)	18.6 (0.1)	90.8 (1.5)	11.7 (1.0)	76.3 (0.9)	6.5 (0.6)

Table 1: NCH linear-probing results (mean (SD) over linear-probe seeds 41–43). Bold/underline: best/second-best. Parentheses under task name: binary-task prevalence. Shaded :

\Delta

F1/

\Delta

AUPRC of best Ours vs. best non-Ours; green = gain.

Model	Modalities	Performance (%)
		Sleep Scoring				Apnea (2.4%)				EEG Arousal (8.0%)				Oxygen Desat. (17.0%)				Hypopnea (3.9%)
		Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR	Acc.	F1	ROC	PR
SleepFM	$c_{1},c_{2}$	70.9 (0.1)	65.5 (0.2)	90.7 (0.1)	68.9 (0.1)	92.7 (0.1)	10.7 (0.1)	73.9 (0.2)	5.4 (0.1)	90.3 (0.1)	52.6 (0.1)	90.6 (0.0)	49.2 (0.1)	81.6 (0.0)	52.3 (0.0)	82.6 (0.1)	49.2 (0.1)	90.4 (0.1)	19.7 (0.0)	79.1 (0.1)	12.4 (0.1)
PedSleepMAE	$c_{1},c_{2}$	44.5 (0.9)	38.1 (0.5)	74.3 (0.1)	39.1 (0.2)	41.4 (5.1)	4.9 (0.2)	60.2 (0.0)	3.0 (0.0)	78.4 (0.8)	27.1 (0.1)	71.2 (0.0)	16.5 (0.0)	67.4 (0.0)	40.9 (0.0)	71.4 (0.0)	33.5 (0.0)	84.3 (2.0)	11.9 (0.2)	67.4 (0.1)	6.1 (0.0)
SynthSleepNet	EEG, EMG	74.9 (0.3)	69.4 (0.3)	92.7 (0.1)	72.8 (0.2)	92.9 (0.6)	13.3 (0.1)	77.3 (0.2)	7.2 (0.0)	89.9 (0.1)	50.6 (0.1)	89.9 (0.0)	46.7 (0.0)	72.5 (0.1)	36.3 (0.1)	69.5 (0.1)	28.4 (0.1)	89.4 (0.4)	14.7 (0.3)	72.1 (0.1)	8.1 (0.1)
NeuroNet	$c_{1}$	67.7 (1.6)	61.5 (0.9)	88.5 (0.4)	64.6 (0.8)	89.9 (0.9)	10.8 (0.3)	72.5 (0.6)	5.2 (0.1)	88.2 (0.8)	44.6 (1.3)	84.8 (0.5)	36.7 (1.2)	50.9 (0.5)	28.8 (0.1)	56.2 (0.0)	21.5 (0.0)	89.4 (0.2)	9.8 (0.8)	69.0 (1.4)	6.0 (0.4)

Table 2: Evaluation of NCH-pretrained models on the external CHAT cohort. Layout and conventions as in Table 1 .

Multimodal Pretraining

Multimodal pretraining learns reusable representations by training models on paired or mixed data from multiple modalities, such as text, images, audio, and signals.

See all 179 papers

Next-Token Prediction Learns Generalisable Representations of Sleep Physiology
Jun 8, 2026Jonathan F. Carter, Lionel TarassenkoSleep Stage Classification Time Series Classification
Institute of Biomedical Engineering University of Oxford
Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS-ANS Dynamics
Jul 4, 2026Zhoujie Hou, Song Wang, Kexin Lou +3Sleep Stage Classification Latent Dynamics Modeling
Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen, China · Omni-Intelligence, Shenzhen, China · Shenzhen Loop Area Institute, Shenzhen, China
Pretraining on Sleep Data Improves non-Sleep Biosignal Tasks
May 4, 2026William Lehn-Schiøler, Magnus Ruud Kjær, Phillip Hempel +7Self-Supervised Time Series Representation Learning Cross-Domain Transfer Learning
BrainCapture ApS, Kongens Lyngby, Denmark · DTU Health Tech, Technical University of Denmark, Kongens Lyngby, Denmark · DTU Compute, Technical University of Denmark, Kongens Lyngby, Denmark +4

Jun 8, 2026J/K move · Enter open · S save

Next-Token Prediction Learns Generalisable Representations of Sleep Physiology

Jonathan F. Carter, Lionel Tarassenko

Institute of Biomedical Engineering University of Oxford

Sleep Stage Classification Time Series Classification Self-Supervised Time Series Representation Learning Multimodal Pretraining Next-Token Prediction Multimodal Foundation Models+2

Open paper AskPDF arXiv

Foundation models offer a promising route to compress multi-modal physiological signals into compact representations of human health, with broad applications across sleep medicine, cardiology, neurology and other healthcare domains. Existing models have typically been trained with masked-reconstruction or contrastive objectives. However, masked reconstruction may be poorly suited to the stochastic nature of these signals, while contrastive approaches rely on positive-pair definitions despite the semantic invariances of physiological signals being poorly understood. In this work, we show that next-token prediction is a simple and scalable alternative. We develop Hypnos, a multi-modal sleep foundation model trained using eight different sensing modalities (e.g. EEG, ECG, respiratory signals) drawn from over 20,000 overnight polysomnography recordings. We tokenize each modality into streams of discrete tokens using residual vector quantization, then train a large auto-regressive RQ-Transformer to jointly predict the next token across all modalities in parallel. After training, Hypnos can be applied to continuous streams of sensor data from any subset of supported modalities, generating embeddings for downstream tasks. Across a range of benchmarks, Hypnos significantly outperforms existing foundation models. In sleep stage classification, we match the performance of strong supervised baselines on held-out test sets whilst using

100\times

less labelled data. Hypnos even generalises to daytime physiology, surpassing a dedicated ECG foundation model at detecting atrial fibrillation. Our results demonstrate that next-token prediction is a strong self-supervised objective for representation learning from multi-modal physiological signals.

Abstract

Multimodal Pretraining

Explore similar work

Next-Token Prediction Learns Generalisable Representations of Sleep Physiology

Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS-ANS Dynamics

Pretraining on Sleep Data Improves non-Sleep Biosignal Tasks