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

Benjamin Wei Hao Chin; Yuin Torng Yew; Jia Hong Puah; Haocheng Wu; Lanxin Liang; Chow Khuen Chan; Norita Mohd Zain; Siti Balqis Samdin

cs.LGAug 21, 2025

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

Authors: Benjamin Wei Hao Chin, Yuin Torng Yew, Jia Hong Puah, Haocheng Wu, Lanxin Liang, Chow Khuen Chan, Norita Mohd Zain, Siti Balqis Samdin,

Organizations: 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.

Abstract

Classification of sleep stages is essential for assessing sleep quality and diagnosing sleep disorders. However, manual inspection of EEG characteristics for each stage is time-consuming and prone to human error. Although machine learning and deep learning methods have been actively developed, they continue to face challenges arising from the non-stationarity and variability of electroencephalography (EEG) and electrooculography (EOG) signals across diverse clinical configurations, often resulting in poor generalization. In this work, we propose SleepDIFFormer, a multi-channel differential transformer framework for heterogeneous EEG-EOG representation learning. SleepDIFFormer is trained across multiple sleep staging datasets, each treated as a source domain, with the goal of generalizing to unseen target domains. Specifically, it employs a Multi-channel Differential Transformer Architecture (MDTA) designed to process raw EEG and EOG signals while incorporating cross-domain alignment. Our approach mitigates spatial and temporal attention noise and learns a domain-invariant EEG-EOG representation through feature distribution alignment across datasets, thereby enhancing generalization to new domains. Empirically, we evaluated SleepDIFFormer on five diverse sleep staging datasets under domain generalization settings and benchmarked it against existing approaches, achieving superior performance. Furthermore, comprehensive ablation analyses and interpretation of the differential attention weights demonstrated their relevance to learning discriminative sleep EEG representations. These findings support the development of more generalizable approaches to automated sleep-stage classification. Our source code and checkpoint are made publicly available at https://github.com/Ben1001409/SleepDIFFormer

Figures & tables

Figure 1: Panel A illustrates the overall architecture of SleepDIFFormer, which is trained on heterogeneous EEG and EOG data from multiple source domains $SD_{\{\cdot\}}$ (collected across different hospitals and recording conditions). To evaluate its cross-domain generalization capability, the trained model is directly tested on a completely unseen target domain $TD_{i}\notin SD_{\{\cdot\}}$ in a strictly zero-shot manner, without any parameter updates or test-time adaptation. Panel B provides a zoomed-in view of the Multi-channel Differential Transformer Architecture (MDTA), which comprises our proposed differential self-attention (DSA) and differential cross-attention (DCA).

Domain $SD_{i}/TD_{i}$	# of Subs	# of Samples	Sampling Freq.	W(%)	N1(%)	N2(%)	N3(%)	REM(%)
1. SleepEDFx	197	236120	100 Hz	29.02	10.64	37.63	8.24	14.47
2. HMC	154	135820	256 Hz	17.00	11.27	36.61	19.59	15.53
3. ISRUC	126	112060	200 Hz	22.01	13.17	31.85	19.69	13.28
4. SHHS-1	5793	132720	125 Hz	30.37	3.19	40.79	13.01	12.64
5. P2018	994	147340	200 Hz	17.36	13.44	42.54	13.27	13.39

Table 1: Summary of Datasets Studied, Characteristics, and Sleep Stage Distribution

Figure 2: Layer-wise differential attention weights for EEG and EOG across different sleep stages. The attention maps are overlaid onto the corresponding waveforms to highlight stage-specific patterns. For visual clarity, the raw EEG and EOG signals were z-score normalized prior to plotting. Consistent color scale is used across all plots.

DA	DSA	DCA	SE	CA	FA	ID	$\mathcal{L}_{\text{rec}}$	$\mathcal{L}_{\text{epo}}$	$\mathcal{L}_{\text{seq}}$	ACC	MF1
✗	✗	✗	✓	✓	✓	✓	✓	✓	✓	75.56	71.44
✓	✗	✓	✓	✓	✓	✓	✓	✓	✓	75.63	71.34
✓	✗	✓	✓	✓	✗	✓	✓	✗	✗	75.71	71.54
✓	✓	✗	✓	✓	✓	✓	✓	✓	✓	76.18	71.55
✓	✓	✗	✓	✓	✗	✓	✓	✗	✗	76.19	71.93
✓	✓	✓	✗	✓	✓	✓	✓	✓	✓	73.40	68.65

Table 2: Ablation study of SleepDIFFormer. Ablating FA is equivalent to ablating both

\mathcal{L}_{\mathrm{epo}}

and

\mathcal{L}_{\mathrm{seq}}

$\lambda_{\mathrm{cls}}$	$\lambda_{\mathrm{rec}}$	$\lambda_{\mathrm{align}}$	Mean ACC	Mean MF1
1.00	0.50	0.50	76.36	72.16
1.00	0.25	0.75	75.76	71.55
1.00	0.75	0.25	75.57	71.46
0.50	0.50	0.50	76.16	71.83
1.00	0.25	0.25	75.73	71.50
Number of MDTA layers			Mean ACC	Mean MF1

Table 3: Sensitivity analysis of hyperparameters in the loss function and MDTA architecture. ACC and MF1 are averaged across five target domains under the leave-one-domain-out protocol. During each analysis, only the specified setting was varied, while all other settings were kept at their default values. The default hyperparameters are underlined, and the best results are highlighted in bold.

	k = 0 (Baseline)		k = 5		k = 10
Dataset	ACC	MF1	ACC	MF1	ACC	MF1
Top-k Occlusion
$\bm{\mathrm{TD}_{1}}$ -SleepEDFx	0.7836	0.7293	0.4463	0.2808	0.3639	0.1906
$\bm{\mathrm{TD}_{2}}$ -HMC	0.7500	0.7271	0.3700	0.3545	0.2172	0.1479
$\bm{\mathrm{TD}_{3}}$ -ISRUC	0.7808	0.7547	0.5205	0.4564	0.3323	0.2334
$\bm{\mathrm{TD}_{4}}$ -SHHS-1	0.7541	0.6759	0.4600	0.3859	0.3500	0.2458

Table 4: Quantitative analysis of differential attention via occlusion experiments. Performance is evaluated when differential attention with the top-

k

, bottom-

k

, near-zero-

k

, random-

k

differential attention weights are occluded. A baseline with no occlusion (

k=0

) is provided for comparison. The best results are highlighted in bold.

	P2018					SHHS-1
	SleepDIFFormer		SleepDG			SleepDIFFormer		SleepDG
Sub. ID	ACC	MF1	ACC	MF1	Sub. ID	ACC	MF1	ACC	MF1
1-150	74.95	72.09	74.13	69.87	1-150	75.41	67.59	73.96	67.12
201-350	73.55	70.79	71.53	67.79	1001-1150	78.29	70.80	75.77	69.22
401-550	71.03	68.13	70.40	65.91	2001-2150	74.95	67.57	74.59	67.79
601-750	72.91	71.17	72.25	69.09	3001-3150	73.70	67.30	72.82	66.78

Table 5: Generalization across different subsets of P2018 and SHHS-1. SleepDIFFormer vs. SleepDG, with the best results highlighted in bold. Rand. (

a

) denotes a randomly selected subset generated using seed

a

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

STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification
May 7, 2026Unaza Tallal, Shruti Kshirsagar, Ankita ShuklaElectroencephalography Sleep Stage Classification
School of Computing, Wichita State University, 1845 N. Fairmount, Wichita, KS 67260, USA · Computer Science and Engineering Department, University of Nevada, Reno, NV, USA
SWINSleepNet: A Hierarchical Context-Aware Framework for Sleep Staging (v2)
Aug 3, 2026Chongjian Wang, Junjie GaoSleep Stage Classification Hybrid CNN-Transformer Architectures
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
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

May 7, 2026J/K move · Enter open · S save

STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification

Unaza Tallal, Shruti Kshirsagar, Ankita Shukla

School of Computing, Wichita State University, 1845 N. Fairmount, Wichita, KS 67260, USA · Computer Science and Engineering Department, University of Nevada, Reno, NV, USA

Electroencephalography Sleep Stage Classification Unsupervised Domain Adaptation Time Series Classification Cross-Dataset Generalization

Open paper AskPDF arXiv

Accurate sleep stage classification across datasets remains challenging due to variability in EEG channel montages, sampling rates, recording environments, and subject populations. Although deep learning has shown considerable promise for automated sleep staging, most existing cross-dataset methods rely on one-dimensional EEG signal representations, whereas the use of two-dimensional spectrogram-based inputs within an unsupervised domain adaptation framework has remained largely unexplored. Here, we propose STDA-Net (Spectrogram-based Temporal Domain Adaptation Network), a framework that combines a convolutional neural network (CNN) for spectrogram-based feature extraction, a bidirectional long short-term memory (BiLSTM) module for temporal modeling of sleep dynamics, and a domain-adversarial neural network (DANN) for source-to-target feature alignment without requiring any labeled target-domain data during training. Experiments are conducted on three publicly available datasets Sleep-EDF, SHHS-1, and SHHS-2 under six cross-dataset transfer settings. Results show that the proposed framework achieves an average accuracy of 89.03% and an average macro F1-score of 87.64%, consistently outperforming existing 1D baseline methods in terms of balanced classification performance, with substantially lower variance across five independent runs, indicating improved stability and reproducibility. Overall, these findings demonstrate that 2D spectrogram-based representations, combined with temporal modeling and adversarial domain adaptation, provide a robust and competitive alternative to conventional 1D EEG inputs for cross-dataset sleep staging.

Abstract

Sleep Stage Classification

Explore similar work

STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification

SWINSleepNet: A Hierarchical Context-Aware Framework for Sleep Staging (v2)

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