[论文解读] ULW-SleepNet: An Ultra-Lightweight Network for Multimodal Sleep Stage Scoring

Zhaowen Wang, Dongdong Zhou+4 more|arXiv (Cornell University)|Feb 27, 2026

EEG and Brain-Computer Interfaces被引用 0

一句话总结

ULW-SleepNet 通过 DSSC 块的超轻量级多模态睡眠分期框架，在 Sleep-EDF 数据集上实现具有竞争力的准确性，同时极大降低参数量与 FLOPs。

ABSTRACT

Automatic sleep stage scoring is crucial for the diagnosis and treatment of sleep disorders. Although deep learning models have advanced the field, many existing models are computationally demanding and designed for single-channel electroencephalography (EEG), limiting their practicality for multimodal polysomnography (PSG) data. To overcome this, we propose ULW-SleepNet, an ultra-lightweight multimodal sleep stage scoring framework that efficiently integrates information from multiple physiological signals. ULW-SleepNet incorporates a novel Dual-Stream Separable Convolution (DSSC) Block, depthwise separable convolutions, channel-wise parameter sharing, and global average pooling to reduce computational overhead while maintaining competitive accuracy. Evaluated on the Sleep-EDF-20 and Sleep-EDF-78 datasets, ULW-SleepNet achieves accuracies of 86.9% and 81.4%, respectively, with only 13.3K parameters and 7.89M FLOPs. Compared to state-of-the-art methods, our model reduces parameters by up to 98.6% with only marginal performance loss, demonstrating its strong potential for real-time sleep monitoring on wearable and IoT devices. The source code for this study is publicly available at https://github.com/wzw999/ULW-SLEEPNET.

研究动机与目标

• Motivate automatic sleep stage scoring for multimodal polysomnography data to support wearable/embedded deployment.
• Develop an ultra-lightweight architecture that scales across multiple input channels without large parameter growth.
• Introduce the Dual-Stream Separable Convolution (DSSC) Block to capture both transient events and long-term temporal features efficiently.
• Reduce computational overhead via depthwise separable convolutions, channel-wise parameter sharing, and global average pooling while preserving accuracy.

提出的方法

• Process multi-channel signals (EEG, EOG, EMG) through a shared channel-wise feature extractor based on DSSC Blocks.
• Use depthwise separable convolutions to minimize parameters and computations.
• Implement channel-wise processing with parameter sharing to handle multiple channels efficiently.
• Incorporate a residual-like DSSC Block architecture to learn hierarchical temporal features.
• Apply global average pooling to replace dense layers and reduce parameters by ~90%.

Fig. 1 : An overview of the ULW-SleepNet architecture. Multimodal physiological input signals (i.e., EEG, EOG, EMG) are processed through a shared channel-wise feature extraction pipeline based on the Dual-Stream Separable Convolutional Block.

实验结果

研究问题

• RQ1 Can an ultra-lightweight multimodal network achieve competitive sleep stage scoring performance on standard PSG datasets?
• RQ2 How do depthwise separable convolutions, channel-wise parameter sharing, and the DSSC Block affect accuracy and efficiency for multimodal sleep staging?
• RQ3 What is the trade-off between model size (parameters) and performance across Sleep-EDF-20 and Sleep-EDF-78?
• RQ4 Is the model suitable for real-time deployment on wearable/embedded hardware without significant accuracy loss?

主要发现

• ULW-SleepNet achieves 86.9% ACC on Sleep-EDF-20 and 81.4% ACC on Sleep-EDF-78 with only 13.3K parameters and 7.89M FLOPs.
• Channel-wise processing with shared parameters improves ACC versus using a single modality.
• Depthwise separable convolutions outperform standard convolutions in this architecture, reducing FLOPs while boosting ACC.
• The model provides strong Wake-stage F1 performance on Sleep-EDF-78 (F1 = 93.0%), while maintaining a small footprint compared to state-of-the-art models.
• Compared to TinySleepNet and LWSleepNet, ULW-SleepNet uses up to 98.6% fewer parameters with substantial FLOPs reduction (up to 85.7%), with only marginal accuracy loss.