Towards Multi-Source Domain Generalization for Sleep Staging with Noisy Labels

Kening Wang; Di Wen; Yufan Chen; Ruiping Liu; Junwei Zheng; Jiale Wei; Kailun Yang; Rainer Stiefelhagen

cs.LGApr 11, 2026

Towards Multi-Source Domain Generalization for Sleep Staging with Noisy Labels

Authors: Kening Wang, Di Wen, Yufan Chen, Ruiping Liu, Junwei Zheng, Jiale Wei, Kailun Yang, Rainer Stiefelhagen,

Organizations: Karlsruhe Institute of Technology · ETH Zurich · Hunan University · INSAIT, Sofia University “St. Kliment Ohridski”

Abstract

Automatic sleep staging is a multimodal learning problem involving heterogeneous physiological signals such as EEG and EOG, which often suffer from domain shifts across institutions, devices, and populations. In practice, these data are also affected by noisy annotations, yet label-noise-robust multi-source domain generalization remains underexplored. We present the first benchmark for Noisy Labels in Multi-Source Domain-Generalized Sleep Staging (NL-DGSS) and show that existing noisy-label learning methods degrade substantially when domain shifts and label noise coexist. To address this challenge, we propose FF-TRUST, a domain-invariant multimodal sleep staging framework with Joint Time-Frequency Early Learning Regularization (JTF-ELR). By jointly exploiting temporal and spectral consistency together with confidence-diversity regularization, FF-TRUST improves robustness under noisy supervision. Experiments on five public datasets demonstrate consistent state-of-the-art performance under diverse symmetric and asymmetric noise settings. The benchmark and code will be made publicly available at https://github.com/KNWang970918/FF-TRUST.git.

Figures & tables

Figure 1. An overview of our task setting, we randomly inject symmetric/asymmetric label noise into the training set according to a predefined noise ratio. On the right-hand side, we deliver the comparison of the Test Accuracy under varying noise conditions, where our approach shows the best performance.

Figure 2. An overview of FF-TRUST architecture, comprising domain-invariant sleep staging feature learning and the proposed joint time–frequency early learning regularization. CE denotes cross-entropy loss, and CDR denotes confidence diversity regularization.

Source Domain $\rightarrow$ Target Domain			II,III,IV,V $\rightarrow$ I		I,III,IV,V $\rightarrow$ II		I,II,IV,V $\rightarrow$ III		I,II,III,V $\rightarrow$ IV		I,II,III,IV $\rightarrow$ V		Average
Method	Setting	NR	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1
SleepDG	BASE	0	76.10	70.88	73.25	70.70	77.38	73.97	75.77	67.70	73.88	69.54	75.28	70.56
	sym.	0.2	75.99	70.68	71.70	69.84	76.36	72.71	75.22	67.40	73.40	68.94	74.53	69.91
		0.4	75.59	69.73	70.40	68.77	75.52	71.79	75.11	66.96	72.64	68.18	73.85	69.09
		0.6	70.43	64.84	65.87	62.09	73.03	66.32	73.04	63.71	70.84	63.36	70.64	64.06
	asym.	0.2	76.14	70.75	73.28	70.89	77.02	73.39	74.40	66.78	73.33	69.45	74.83	70.25

Table 1. Performance comparison of FF-TRUST with existing LNL methods under symmetric and asymmetric label noise. NR : noise rate, I : Sleep-EDFx, II : HMC, III : ISRUC, IV : SHHS, V : P2018.

Source Domain $\rightarrow$ Target Domain			II,III,IV,V $\rightarrow$ I		I,III,IV,V $\rightarrow$ II		I,II,IV,V $\rightarrow$ III		I,II,III,V $\rightarrow$ IV		I,II,III,IV $\rightarrow$ V		Average
Method	Setting	NR	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1
DeepSleepNet	BASE	0	74.10	68.69	70.28	67.75	77.47	73.65	74.62	67.22	71.23	66.66	73.54	68.79
	sym.	0.2	72.68	67.79	69.77	67.37	77.01	72.80	73.49	65.89	70.80	66.75	72.75	68.12
		0.4	70.69	65.44	67.16	64.85	76.32	71.28	74.19	66.33	70.74	65.39	71.82	66.66
		0.6	69.60	62.37	65.37	61.71	74.40	67.94	73.27	64.68	69.23	61.71	70.37	63.68
	asym.	0.2	73.07	67.14	68.69	66.58	76.96	72.49	73.78	66.16	70.90	66.46	72.68	67.77

Table 2. Performance comparison of FF-TRUST with existing model under symmetric and asymmetric label noise. NR : noise rate, I : Sleep-EDFx, II : HMC, III : ISRUC, IV : SHHS, V : P2018.

Modules			II,III,IV,V $\rightarrow$ I		I,III,IV,V $\rightarrow$ II		I,II,IV,V $\rightarrow$ III		I,II,III,V $\rightarrow$ IV		I,II,III,IV $\rightarrow$ V		Average
TimeELR	FourierELR	FF-CDR	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1
			70.43	64.84	65.87	62.09	73.03	66.32	73.04	63.71	70.84	63.36	70.64	64.06
✓			71.54	65.20	68.37	66.15	73.21	68.73	72.01	65.80	70.82	65.63	71.19	66.30
		✓	71.44	65.83	68.41	66.95	73.37	69.50	74.18	65.74	70.62	66.25	71.60	66.85
	✓		71.63	65.08	68.46	66.09	73.76	69.14	74.23	65.44	71.66	66.38	71.95	66.43
	✓	✓	71.54	66.09	66.23	63.72	73.84	69.93	73.63	65.59	71.81	67.39	71.41	66.54

Table 3. Module ablation study under

0.6

symmetric label noise. We verify the effectiveness of TimeELR, FourierELR and FF-CDR.

Panel of Figure 3 — Figure 3. Comparison of hypnograms generated by the baseline model (a) DISC and (b) the proposed FF-TRUST model on Sleep-EDFx under $NR=0.6$ symmetric label noise. True sleep stages (blue), predicted stages (red dashed) and mismatches (purple bar at bottom) are visualized over time.

Panel of Figure 4 — Figure 4. T-SNE visualizations of feature distributions for DISC and FF-TRUST on the HMC and ISRUC datasets. (a1), (a2) show DISC on HMC and ISRUC, respectively, while (b1), (b2) show FF-TRUST .

Panel of Figure 5 — Figure 5. Comparison of training accuracy (a) and MF1 score (b) over epochs for FF-TRUST and SleepDG on the HMC dataset under $NR=0.6$ asymmetric label noise.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Source Domain $\rightarrow$ Target Domain			II,III,IV,V $\rightarrow$ I		I,III,IV,V $\rightarrow$ II		I,II,IV,V $\rightarrow$ III		I,II,III,V $\rightarrow$ IV		I,II,III,IV $\rightarrow$ V		Average
Method	Setting	NR	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1	ACC	MF1
FourierELR(log)	sym.	0.2	76.83	71.63	72.06	70.19	77.15	74.28	76.59	68.02	73.89	70.44	75.30	70.91
		0.4	76.30	70.98	71.14	68.86	76.40	73.16	76.07	67.84	72.48	68.40	74.48	69.85
		0.6	72.80	66.00	68.38	64.72	73.78	68.75	75.34	66.01	71.96	66.41	72.45	66.38
	asym.	0.2	77.05	71.26	73.54	71.28	77.07	73.42	75.59	67.60	73.85	69.48	75.42	70.61
		0.4	76.44	71.09	72.56	70.84	76.53	72.77	75.80	67.60	73.79	69.49	75.02	70.36

Appendix

Table 4. Performance comparison of FourierELR with variant under symmetric and asymmetric label noise. NR : noise rate, I : Sleep-EDFx, II : HMC, III : ISRUC, IV : SHHS, V : P2018.

Panel of Figure 6 — Figure 6. Comparison of hypnograms generated by the baseline model (a) DISC and (b) the proposed FF-TRUST model on HMC under $NR=0.6$ symmetric label noise. True sleep stages (blue) and predicted stages (red dashed) are visualized over time.

Panel of Figure 7 — Figure 7. Comparison of hypnograms generated by the baseline model (a) DISC and (b) the proposed FF-TRUST model on ISRUC under $NR=0.6$ symmetric label noise. True sleep stages (blue) and predicted stages (red dashed) are visualized over time.

Panel of Figure 8 — Figure 8. Comparison of hypnograms generated by the baseline model (a) DISC and (b) the proposed FF-TRUST model on SHHS under $NR=0.6$ symmetric label noise. True sleep stages (blue) and predicted stages (red dashed) are visualized over time.

Panel of Figure 9 — Figure 9. Comparison of hypnograms generated by the baseline model (a) DISC and (b) the proposed FF-TRUST model on P2018 under $NR=0.6$ symmetric label noise. True sleep stages (blue) and predicted stages (red dashed) are visualized over time.

Source Domain

\rightarrow

Target Domain

II,III,IV,V

\rightarrow

I,III,IV,V

\rightarrow

I,II,IV,V

\rightarrow

III

I,II,III,V

\rightarrow

I,II,III,IV

\rightarrow

Average

Method

Setting

ACC

MF1

ACC

MF1

ACC

MF1

ACC

MF1

ACC

MF1

ACC

MF1

DISC

sym.

0.6

71.51\pm 0.49

64.87\pm 0.53

67.14\pm 0.98

64.19\pm 0.96

73.20\pm 0.51

67.84\pm 1.12

73.45\pm 1.41

64.88\pm 1.16

70.76\pm 1.56

64.81\pm 1.65

71.21

65.32

asym.

0.6

67.34\pm 2.09

60.77\pm 2.18

59.77\pm 3.01

56.28\pm 3.78

64.35\pm 2.15

59.09\pm 2.74

67.53\pm 1.08

59.21\pm 1.72

66.03\pm 1.18

63.80\pm 1.33

65.00

59.83

FF-TRUST(ours)

sym.

0.6

73.57\pm 1.35

67.51\pm 1.22

69.64\pm 0.82

67.76\pm 0.81

74.42\pm 0.38

70.26\pm 0.65

74.59\pm 0.93

65.58\pm 0.29

71.89\pm 0.85

68.13\pm 0.65

72.82

67.85

asym.

0.6

71.25\pm 0.82

66.34\pm 0.47

68.60\pm 0.53

66.18\pm 0.74

72.35\pm 0.25

68.79\pm 0.49

72.61\pm 0.94

64.74\pm 0.43

71.24\pm 0.77

69.07\pm 0.37

71.21

67.02

Appendix

Table 5. Performance comparison of FF-TRUST with DISC under NR=0.6 symmetric and asymmetric label noise across multiple random seeds. NR : noise rate, I : Sleep-EDFx, II : HMC, III : ISRUC, IV : SHHS, V : P2018.

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

SleepBand: Single-Source Domain Generalization for Sleep Staging via Physiologically Structured Spectral Modeling
Jul 6, 2026Zhi Lu, Yang Hu, Yan ChenSleep Stage Classification EEG Decoding
Laboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China, Chengdu 611731, China · School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China · School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China
DualStabSleepNet: A Dual-Domain Diffusion Stabilization Network for Robust Sleep Staging
Aug 17, 2026Chongjian Wang, Chen Liu, Junjie Gao +3Diffusion-Based Denoising Sleep Stage Classification
School of Mathematics and Systems Science, Shandong University of Science and Technology, Qingdao, 266590, Shandong, China · School of Artificial Intelligence, Shandong Women’s University, Jinan, 250352, Shandong, China · School of Computer Science and Engineering, South China University of Technology, Guangzhou, 510641, Guangdong, China
Sleep Stage Classification across Heterogeneous EEG/EOG Channels with Multi-Channel Differential Transformer
Aug 21, 2025Benjamin Wei Hao Chin, Yuin Torng Yew, Jia Hong Puah +6Sleep Stage Classification Transformer
School of Artificial Intelligence and Robotics, Xiamen University Malaysia. · Department of Biomedical Engineering, Universiti Malaya. · Department of Biomedical Engineering and Health Sciences, Universiti Teknologi Malaysia.

Jul 6, 2026J/K move · Enter open · S save

SleepBand: Single-Source Domain Generalization for Sleep Staging via Physiologically Structured Spectral Modeling

Zhi Lu, Yang Hu, Yan Chen

Laboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China, Chengdu 611731, China · School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China · School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China

Sleep Stage Classification EEG Decoding Frequency-Domain Feature Learning Domain Generalization Cross-Dataset Generalization

Open paper AskPDF arXiv

Generalizing sleep staging models to unseen datasets is challenging, and typical domain generalization (DG) methods often rely on multiple source domains or domain labels that are rarely available in practice. We tackle the stricter and more practical setting of single-source domain generalization: training on a single labeled source dataset, without domain labels or access to target data. We present SleepBand, a physiology-guided framework that embeds oscillatory priors via a learnable Morlet filter bank and a structured integration-and-recalibration pipeline. This anchors representations to domain-invariant sleep rhythms (e.g., slow waves, spindles), reducing reliance on dataset-specific artefacts. On five public datasets, SleepBand achieves state-of-the-art SDG performance and remains competitive under leave-one-domain-out (multi-source) DG. Analyses show that the learned filters align with canonical neurophysiology and that robustness stems from focusing on narrowband, physiologically meaningful cues. Our results suggest that principled, physiology-aware inductive biases are a promising path for robust single-domain sleep staging. Code is available at https://github.com/lzcn/sleep-band

Abstract

Sleep Stage Classification

Explore similar work

SleepBand: Single-Source Domain Generalization for Sleep Staging via Physiologically Structured Spectral Modeling

DualStabSleepNet: A Dual-Domain Diffusion Stabilization Network for Robust Sleep Staging

Sleep Stage Classification across Heterogeneous EEG/EOG Channels with Multi-Channel Differential Transformer