Combining scEEG and PPG for reliable sleep staging

Jiawei Wang; Liang Xu; Shuntian Zheng; Yu Guan; Kaichen Wang; Ziqing Zhang; Chen Chen; Laurence T. Yang

eess.SPFeb 4, 2026

Combining scEEG and PPG for reliable sleep staging

Authors: Jiawei Wang, Liang Xu, Shuntian Zheng, Yu Guan, Kaichen Wang, Ziqing Zhang, Chen Chen, Laurence T. Yang,

Organizations: School of Engineering, The University of Warwick, Coventry, UK · Department of Computer Science, The University of Warwick, Coventry, UK · College of Biomedical Engineering, Fudan University, Shanghai, China · Human Phenome Institute, Fudan University, Shanghai, China ·

Abstract

Reliable sleep staging remains challenging for lightweight wearable devices such as single-channel electroencephalography (scEEG) or photoplethysmography (PPG). scEEG offers direct measurement of cortical activity and serves as the foundation for sleep staging, yet exhibits limited performance on light sleep stages. PPG provides a low-cost complement that captures autonomic signatures effective for detecting light sleep. However, prior PPG-based methods rely on full night recordings (8 - 10 hours) as input context, which is less practical to provide timely feedback for sleep intervention. In this work, we investigate scEEG-PPG fusion for 4-class sleep staging under short-window (30 s - 30 min) constraints. First, we evaluate the temporal context required for each modality, to better understand the relationship of sleep staging performance with respect to monitoring window. Second, we investigate three fusion strategies: score-level fusion, cross-attention fusion enabling feature-level interactions, and Mamba-enhanced fusion incorporating temporal context modeling. Third, we train and evaluate on the Multi-Ethnic Study of Atherosclerosis (MESA) dataset and perform cross-dataset validation on the Cleveland Family Study (CFS) and the Apnea, Bariatric surgery, and CPAP (ABC) datasets. The Mamba-enhanced fusion achieves the best performance on MESA (Cohen's Kappa $κ$ = 0.798, Acc = 86.9%), with particularly notable improvement in light sleep classification (F1-score: 85.63% vs. 77.76%, recall: 82.85% vs. 69.95% for scEEG alone), and generalizes well to CFS and ABC datasets with different populations. These findings suggest that scEEG-PPG fusion is a promising approach for lightweight wearable based sleep monitoring, offering a pathway toward more accessible sleep health assessment. Source code of this project can be found at: https://github.com/DavyWJW/scEEG-PPGFusion

Figures & tables

Fig. 1 : Wearable sleep monitoring devices: (a) clinical gold-standard PSG [ 4 ] , (b) multi-signal system integrating EEG, EOG, and EMG (Tosoo AG) [ 7 ] , (c) multi-channel EEG system (Greentek GT Cap) [ 10 ] , (d) forehead scEEG device (Neurovista Sleep Monitor) [ 11 ] , (e) headset scEEG device (Brainlink Lite) [ 12 ] , (f) PPG-based smart ring (RingConn) [ 13 ] , (g) PPG-based smart watch (Garmin) [ 14 ] .

Fig. 2 : Demographic characteristics of MESA, CFS and ABC datasets: (a) Age. (b) Sex distribution. (c) Racial composition highlighting MESA’s ethnic diversity, CFS’s predominantly Black and White population, and ABC’s majority White cohort.

Fig. 3 : Architecture of baseline models: (a) scEEG model, (b) PPG model.

Fig. 4 : Overview of three fusion strategies: (a) score-level fusion, (b) cross-attention fusion, and (c) Mamba-enhanced fusion.

Fig. 5 : Confusion matrices for sleep stage classification across different window lengths using scEEG and PPG.

Window	$\kappa$	Acc	Model Size	Infer. (ms)
30 s	0.665	0.772	0.52M	3.84 $\pm$ 0.23
1 min	0.674	0.783	2.71M	3.92 $\pm$ 0.24
3 min	0.697	0.799	2.71M	3.97 $\pm$ 0.18
5 min	0.721	0.815	2.71M	3.98 $\pm$ 0.26
10 min	0.688	0.787	2.71M	3.94 $\pm$ 0.27
30 min	0.735	0.821	2.71M	3.94 $\pm$ 0.32

TABLE I : scEEG model performance across window lengths on MESA test set (M=204).

Window	$\kappa$	Acc	Model Size	Infer.(ms)
30 s	0.460	0.647	8.28M	8.44 $\pm$ 0.56
1 min	0.538	0.701	8.29M	9.10 $\pm$ 0.68
3 min	0.643	0.772	10.26M	9.39 $\pm$ 0.39
5 min	0.633	0.758	10.26M	9.41 $\pm$ 0.55
10 min	0.656	0.777	12.23M	10.28 $\pm$ 0.59
30 min	0.696	0.804	14.20M	11.05 $\pm$ 0.60

TABLE II : PPG model performance across window lengths on the MESA test set (M=204).

Strategy	Method	$\boldsymbol{\kappa}$	Acc	Model Size	Infer. (ms)
PPG	SleepPPGNet [ 18 ]	0.479	0.659	2.29M	8.10 $\pm$ 0.35
PPG	Dual-stream PPG [ 22 ]	0.643	0.772	10.26M	9.39 $\pm$ 0.39
scEEG	AttnSleep [ 19 ]	0.697	0.799	2.71M	3.97 $\pm$ 0.18
scEEG	CBraMod [ 37 ]	0.762	0.769	12.14M	6.18 $\pm$ 0.79
Fusion	Score-level (ours)	0.747	0.835	12.9M	14.40 $\pm$ 1.07
	Cross-attention (ours)	0.767	0.844	16.29M	15.69 $\pm$ 1.27

TABLE III : Comparison of fusion models and SOTA methods under the 3-minute window on the MESA dataset.

Fig. 6 : Per-subject Cohen’s

\kappa

, light-sleep F1, and recall on MESA (

M\!=\!204

); ∗∗∗ :

p<0.001

(Wilcoxon) for Mamba-enhanced fusion vs. each single modality.

Temporal module	Train. params	$\boldsymbol{\kappa}$ (95% CI)
Transformer	5.21 M	0.786 [0.775, 0.798]
Bi-GRU	5.20 M	0.783 [0.774, 0.793]
Bi-LSTM	5.22 M	0.781 [0.772, 0.791]
Bi-Mamba	5.21 M	0.787 [0.777, 0.798]

TABLE IV : Parameter-matched temporal-module ablation on the MESA test set, where only the TCM differs with trainable parameters matched; per-subject

\kappa

is mean [95% CI].

Fig. 7 : Effect of fusion weight $\alpha$ on performance under the 3-minute window.

Fig. 8 : Sleep architecture estimation errors (MAE) on MESA test set.

Fig. 9 : Sleep-stage transition-probability matrices: ground truth (left) vs. Mamba-enhanced fusion (right) on MESA.

Factor	Subgroup	n	$\boldsymbol{\kappa}$ (95% CI)	Acc (%)
OSA (AHI)	Normal ( $<5$ )	17	0.828 [0.795, 0.853]	89.1
	Mild ( $5$ – $15$ )	64	0.786 [0.765, 0.806]	86.6
	Moderate ( $15$ – $30$ )	66	0.788 [0.769, 0.807]	86.7
	Severe ( $\geq 30$ )	57	0.775 [0.757, 0.791]	86.5
Age (yr)	$<65$	67	0.787 [0.769, 0.805]	86.6
	$65$ – $74$	67	0.799 [0.781, 0.817]	87.7

TABLE V : Subgroup analysis of the Mamba-enhanced fusion on the MESA test set (

n\!=\!204

): per-subject

\kappa

(mean [95% CI]) and accuracy.

Fig. 10 : Synchronization sensitivity:

\kappa

and light-sleep F1 of the Mamba-enhanced fusion versus an imposed scEEG-PPG temporal offset

\Delta

(MESA).

Strategy	CFS		ABC
Strategy	$\kappa$	Acc	$\kappa$	Acc
PPG (w/o fine-tuning)	0.090	0.432	0.368	0.66
scEEG (w/o fine-tuning)	0.638	0.741	0.334	0.535
PPG (fine-tuning)	0.583	0.710	0.494	0.691
scEEG (fine-tuning)	0.716	0.799	0.543	0.681

TABLE VI : Single-modality generalization performance across datasets.

Dataset	Fusion Method	$\boldsymbol{\kappa}$	Acc	Model Size	Infer. (ms)
CFS	Score-level	0.660	0.772	12.9M	14.62 $\pm$ 1.25
	Cross-attention	0.781	0.849	16.29M	16.05 $\pm$ 1.14
	Mamba-enhanced	0.769	0.839	18.17M	19.20 $\pm$ 1.47
	Cross-attention (zs)	0.235	0.506	16.29M	16.05 $\pm$ 1.14
	Mamba-enhanced (zs)	0.152	0.462	18.17M	19.20 $\pm$ 1.47
ABC	Score-level	0.524	0.684	12.9M	13.38 $\pm$ 0.56

TABLE VII : Cross-dataset generalization of fusion methods. Rows are fine-tuned unless marked zero-shot (zs; direct transfer with no target-domain training.)

Fig. 11 : Comparison of $\kappa$ across single-modality and fusion-based methods on the MESA, CFS, and ABC datasets.

Fig. 12 : Confusion matrix of the Mamba-enhanced fusion model on the MESA dataset.

Fig. 13 : Fusion interpretability (MESA): (a) t-SNE of the fused features by stage; (b) per-stage F1.

Sleep Stage Classification

Sleep stage classification assigns sleep epochs to physiological stages from signals such as EEG, helping quantify sleep architecture and support clinical monitoring.

See all 40 papers

Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks
Aug 2, 2026Shuntian Zheng, Jiawei Wang, Cong Fu +4Sleep Stage Classification Photoplethysmography
Department of Computer Science, The University of Warwick · School of Engineering, The University of Warwick · Department of Neurology, Huashan Hospital, Fudan University +2
Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction
Aug 14, 2026Hassan Mehdi, Riku Klen, Ayse Kosal Bulbul +5Sleep Stage Classification EEG Foundation Models
Institute of Biomedicine, University of Turku, Finland · Turku PET Centre, University of Turku and Turku University Hospital, Finland · University College Cork, Ireland +3
Personalized Scorer Modeling: A Learning-Based Framework for Deriving Robust Sleep Stage Labels from Multiple Experts
Aug 12, 2026Seyyed Ali Hoseini, Javad Baseri, Hamid Saadatfar +2Sleep Stage Classification Human-in-the-Loop Annotation
Department of Electrical and Computer Engineering, University of Birjand, Birjand, South Khorasan, Iran

Aug 2, 2026J/K move · Enter open · S save

Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks

Shuntian Zheng, Jiawei Wang, Cong Fu, Huan Yu+3

Department of Computer Science, The University of Warwick · School of Engineering, The University of Warwick · Department of Neurology, Huashan Hospital, Fudan University+2

Sleep Stage Classification Photoplethysmography

Open paper AskPDF arXiv

Automated sleep staging assigns discrete stage labels to successive time epochs throughout an overnight recording; conventionally each window spans at least 30 seconds, reflecting the minimum temporal resolution of the clinical scoring standard. Wearable photoplethysmography (PPG) has attracted sustained interest as an ambulatory alternative to laboratory-based polysomnography, which relies on electroencephalography (EEG) and other recording modalities that are impractical outside clinical environments. Yet PPG-based staging trails EEG-based methods by a substantial margin, and we argue this gap largely reflects a mismatch between signal and task. Within a stable stage, PPG's inter-stage feature differences are more subtle than those in EEG; yet at stage boundaries, PPG's principal cardiovascular features, heart rate variability and pulse morphology, shift sharply within seconds. The conventional practice of assigning one label to each 30-second epoch therefore suppresses feature that is concentrated near boundaries. We address this gap in two steps. First, we develop a label expansion pipeline based on Hidden Semi-Markov Models that converts coarse epoch labels into sec-level annotations. To assess whether these expanded labels are reliable enough for downstream supervision, we validate them on a separate expert-reviewed dataset and through an auxiliary sleep-wake task whose labels are independent of the expansion pipeline. Second, we use the resulting sec-level supervision on MESA to improve conventional four-class epoch-level staging across four architecturally diverse baselines by 3.7--5.7,pp in accuracy against the original epoch labels, with supplementary zero-shot evaluation on CFS showing that the transfer benefit persists under cohort and annotation-protocol shift.

Abstract

Sleep Stage Classification

Explore similar work

Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks

Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction

Personalized Scorer Modeling: A Learning-Based Framework for Deriving Robust Sleep Stage Labels from Multiple Experts