NanoSleep: A Parameter-Efficient Hybrid Temporal Convolutional Network for Single-Channel Sleep Stage Classification

S M Asif Hossain; Shruti Kshirsagar

cs.LGAug 19, 2026

NanoSleep: A Parameter-Efficient Hybrid Temporal Convolutional Network for Single-Channel Sleep Stage Classification

Authors: S M Asif Hossain, Shruti Kshirsagar

Organizations: School of Computing, Wichita State University, Kansas, USA

Abstract

Sleep stage classification from single-channel electroencephalography (EEG) is essential for wearable and home-based sleep monitoring. However, many deep learning models achieve high accuracy at the cost of large model sizes, which limits their deployment on resource-constrained devices. In this work, we present NanoSleep, a compact hybrid temporal convolutional network for automatic sleep stage classification. NanoSleep combines a learnable Sinc-convolutional front end, a dual-branch feature extractor that fuses multi-scale temporal and spectral representations, a gated dilated temporal convolutional backbone with channel recalibration, and a conditional random field for sequence-level decoding. We further employ a weighted calibrated focal loss to address class imbalance. We evaluate NanoSleep on the Sleep-EDF and Sleep-EDF-Expanded datasets using subject-wise cross-validation. The proposed model consistently outperforms six representative baseline methods, and an ablation study confirms the contribution of each major component. These results demonstrate that NanoSleep provides an effective balance between accuracy and efficiency, making it well suited for wearable devices, home-based sleep monitoring, and resource-constrained clinical applications.

Figures & tables

Figure 1: Overall architecture of the proposed NanoSleep framework for single-channel EEG sleep stage classification.

Figure 2: Power spectral density of the single-channel EEG for the five sleep stages, with the standard delta, theta, alpha, and beta frequency bands indicated.

Property	Sleep-EDF	Sleep-EDF-Exp.
Subjects	20	78
EEG channel	Fpz-Cz	Fpz-Cz
Sampling rate	100 Hz	100 Hz

Table 1: Composition of the two datasets used in this study.

Figure 3: Representative 30 s single-channel EEG epochs from the Fpz-Cz channel of the Sleep-EDF dataset for the five sleep stages. The N2 epoch exhibits sleep spindles and K-complexes, the N3 epoch is dominated by high-amplitude slow waves, and the REM epoch presents a low-amplitude mixed-frequency pattern.

Figure 4: Full-night hypnogram of a representative subject from the Sleep-EDF dataset, illustrating the cyclic alternation of sleep stages across the night and the comparatively small proportion of the N1 stage.

Figure 5: Effect of the data augmentation operations on a representative N2 epoch. Panel (a) shows the original epoch, panel (b) shows the epoch after brain-wave stretching, and panel (c) shows the epoch after brain-noise injection with a slow voltage drift.

Model	ACC (%)	MF1 (%)	$\kappa$	W	N1	N2	N3	REM	Param.
SleepEEGNet	84.3 $\pm$ 0.4	79.7 $\pm$ 0.8	0.79 $\pm$ 0.02	89.2 $\pm$ 0.5	52.2 $\pm$ 0.8	86.8 $\pm$ 0.3	85.1 $\pm$ 0.3	85.0 $\pm$ 0.3	2.6 M
DeepSleepNet	81.9 $\pm$ 0.3	76.6 $\pm$ 0.3	0.76 $\pm$ 0.04	86.7 $\pm$ 0.4	45.5 $\pm$ 0.3	85.1 $\pm$ 0.3	83.3 $\pm$ 0.4	82.6 $\pm$ 0.2	24.7 M
XSleepNet	83.9 $\pm$ 0.3	78.7 $\pm$ 0.2	0.77 $\pm$ 0.04	81.6 $\pm$ 0.4	52.9 $\pm$ 0.3	88.1 $\pm$ 0.2	85.3 $\pm$ 0.3	85.4 $\pm$ 0.3	5.8 M
AttnSleep	84.4 $\pm$ 0.3	78.1 $\pm$ 0.3	0.79 $\pm$ 0.03	89.7 $\pm$ 0.3	42.6 $\pm$ 0.3	88.8 $\pm$ 0.3	90.2 $\pm$ 0.3	79.0 $\pm$ 0.4	0.52 M
DeepSleepNet-Lite	84.0 $\pm$ 0.4	78.0 $\pm$ 0.3	0.78 $\pm$ 0.04	87.1 $\pm$ 0.4	44.4 $\pm$ 0.3	87.9 $\pm$ 0.2	88.2 $\pm$ 0.4	82.4 $\pm$ 0.3	0.6 M
TinySleepNet	85.4 $\pm$ 0.3	80.5 $\pm$ 0.3	0.80 $\pm$ 0.04	90.1 $\pm$ 0.3	51.4 $\pm$ 0.3	88.5 $\pm$ 0.2	88.3 $\pm$ 0.3	84.3 $\pm$ 0.4	1.3 M

Table 2: Performance comparison on the Sleep-EDF dataset under 20-fold subject-wise cross-validation.

Model	ACC (%)	MF1 (%)	$\kappa$	W	N1	N2	N3	REM	Param.
SleepEEGNet	80.0 $\pm$ 0.5	73.6 $\pm$ 0.5	0.73 $\pm$ 0.03	91.7 $\pm$ 0.5	44.1 $\pm$ 1.7	82.5 $\pm$ 0.6	73.5 $\pm$ 1.0	76.1 $\pm$ 1.0	2.6 M
DeepSleepNet	77.8 $\pm$ 0.5	71.8 $\pm$ 0.6	0.70 $\pm$ 0.03	90.9 $\pm$ 0.7	45.0 $\pm$ 2.2	79.2 $\pm$ 0.6	72.7 $\pm$ 0.9	71.1 $\pm$ 0.7	24.7 M
XSleepNet	80.3 $\pm$ 0.5	76.4 $\pm$ 0.5	0.72 $\pm$ 0.02	85.2 $\pm$ 0.5	49.4 $\pm$ 1.4	86.0 $\pm$ 0.6	79.8 $\pm$ 0.8	81.7 $\pm$ 0.9	5.8 M
AttnSleep	81.3 $\pm$ 0.5	75.2 $\pm$ 0.7	0.74 $\pm$ 0.03	92.0 $\pm$ 0.7	42.9 $\pm$ 1.0	85.0 $\pm$ 0.7	82.1 $\pm$ 0.6	74.2 $\pm$ 0.9	0.52 M
DeepSleepNet-Lite	80.3 $\pm$ 0.5	75.2 $\pm$ 0.7	0.73 $\pm$ 0.02	91.5 $\pm$ 0.8	46.4 $\pm$ 1.0	82.9 $\pm$ 0.7	79.2 $\pm$ 0.8	76.4 $\pm$ 0.5	0.6 M
TinySleepNet	83.1 $\pm$ 0.4	78.2 $\pm$ 0.5	0.77 $\pm$ 0.02	92.8 $\pm$ 0.7	51.5 $\pm$ 1.9	85.3 $\pm$ 0.7	81.1 $\pm$ 0.8	80.3 $\pm$ 0.9	1.3 M

Table 3: Performance comparison on the Sleep-EDF-Expanded dataset under 10-fold subject-wise cross-validation.

Panel of Figure 6 — Figure 6: Confusion matrices of NanoSleep on the two evaluation datasets. Rows denote the expert annotation and columns denote the model prediction.

Model Variant	ACC (%)	MF1 (%)	$\kappa$	W	N1	N2	N3	REM
W/o spectral branch	83.5 $\pm$ 0.4	77.9 $\pm$ 0.5	0.76 $\pm$ 0.03	92.6 $\pm$ 0.7	43.5 $\pm$ 1.5	86.2 $\pm$ 0.4	84.5 $\pm$ 0.4	83.0 $\pm$ 0.7
W/o multi-scale convolution (single branch)	82.0 $\pm$ 0.4	76.7 $\pm$ 0.6	0.74 $\pm$ 0.03	88.8 $\pm$ 0.5	42.2 $\pm$ 1.6	85.1 $\pm$ 0.5	83.2 $\pm$ 0.7	84.2 $\pm$ 0.5
W/o conditional random field (softmax decoding)	84.8 $\pm$ 0.5	79.9 $\pm$ 0.6	0.79 $\pm$ 0.03	92.5 $\pm$ 0.9	47.3 $\pm$ 1.3	88.5 $\pm$ 0.4	86.0 $\pm$ 0.6	85.5 $\pm$ 0.5
W/o squeeze-and-excitation block	85.8 $\pm$ 0.5	81.3 $\pm$ 0.5	0.80 $\pm$ 0.04	94.1 $\pm$ 0.6	49.2 $\pm$ 1.5	89.2 $\pm$ 0.4	87.8 $\pm$ 0.8	86.3 $\pm$ 0.7
W/o focal loss (standard cross-entropy)	85.2 $\pm$ 0.4	79.5 $\pm$ 0.6	0.79 $\pm$ 0.04	94.8 $\pm$ 0.4	44.1 $\pm$ 1.7	89.0 $\pm$ 0.5	85.2 $\pm$ 0.7	84.5 $\pm$ 0.7
W/o data augmentation	84.5 $\pm$ 0.4	79.8 $\pm$ 0.6	0.78 $\pm$ 0.02	92.8 $\pm$ 0.5	47.0 $\pm$ 1.7	88.2 $\pm$ 0.6	86.1 $\pm$ 0.5	85.0 $\pm$ 0.6

Table 4: Ablation study of the proposed NanoSleep model on the Sleep-EDF dataset under 20-fold subject-wise cross-validation.

Explore similar work

Lightweight ML-Based Automatic Sleep Staging Framework with Constrained CNN and Mamba for Small-Sample EEG Datasets
Jul 6, 2026Zihao Wei, Yulin Gong, Yudan LvEfficient Neural Network Inference Electroencephalography
School of Electronic Information Engineering, Changchun University of Science and Technology, Changchun 130022, Jilin Province, China · Jilin University, Changchun 130012, Jilin Province, China
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
Energy-Efficient Real-Time 4-Stage Sleep Classification at 10-Second Resolution
Aug 6, 2025Zahra Mohammadi, Parnian Fazel, Siamak MohammadiSleep Stage Classification Energy-Efficient ML
School of Electrical and Computer Engineering, University of Tehran.

Jul 6, 2026, cs.LGJ/K move · Enter open · S save

Lightweight ML-Based Automatic Sleep Staging Framework with Constrained CNN and Mamba for Small-Sample EEG Datasets

Zihao Wei, Yulin Gong, Yudan Lv

School of Electronic Information Engineering, Changchun University of Science and Technology, Changchun 130022, Jilin Province, China · Jilin University, Changchun 130012, Jilin Province, China

Efficient Neural Network Inference Electroencephalography Sleep Stage Classification Data-Efficient Learning

Open paper AskPDF arXiv

Automatic sleep staging is a key technology for precise diagnosis and treatment of sleep disorders as well as long-term home sleep monitoring. Portable electroencephalogram (EEG) devices have become the focus of research due to their convenience in data collection. However, current methods still face three major challenges: large parameter sizes that easily lead to overfitting on small datasets, low accuracy in classifying difficult stages such as N1 and REM, unclear optimal training dataset size, and difficulty in deployment. This paper proposes GamSleepNet, a lightweight and low-latency automatic sleep staging framework for single-channel EEG. The framework features the FEB module, which combines improved Gabor kernels with learnable filters for feature extraction, uses the Mamba architecture to build a temporal classification network, introduces a novel contrastive loss and a two-stage training strategy, and experimentally validates the optimal dataset size for single-channel EEG sleep staging models. On the Sleepedf dataset, this model achieves an overall accuracy of 87.86 percent with only 30.86 thousand parameters, with all metrics reaching SOTA levels and significantly improving the identification accuracy of challenging sleep stages.

Abstract

Explore similar work

Lightweight ML-Based Automatic Sleep Staging Framework with Constrained CNN and Mamba for Small-Sample EEG Datasets

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

Energy-Efficient Real-Time 4-Stage Sleep Classification at 10-Second Resolution