Source-linked AI summary
LightSleepX: A Lightweight, Inception-Based Dual-Modal Network for Sleep Staging
Yi Wang
TL;DR
Manual PSG scoring is cumbersome for home tracking, while deep models can be costly and shallow methods may miss sleep-signal dynamics. LightSleepX combines multi-modal feature extraction with Mamba temporal modeling, reporting strong benchmark performance with a small parameter count. Its stated target is practical, privacy-preserving local sleep monitoring.
Problem
Manual PSG scoring is cumbersome for continuous home tracking, and existing deep models can be computationally demanding while shallow methods may fail to capture sleep's temporal dynamics.
Source: §1 · p. 1Method
LightSleepX combines an Inception-style EEG/EOG feature extractor with Multi-scale Enhanced Attention and a Mamba encoder for rule-free long-range temporal modeling.
Source: §1 · p. 1, §2.2.4 · p. 3Results
LightSleepX reports top performance across Sleep-EDF-20 metrics and the highest Macro-F1 on Sleep-EDF-78, with 81.80% accuracy on cross-subject ISRUC-S3.
Source: §3.2 · p. 5Takeaways & Limitations
Abstract
from arXiv · showhide
Automatic sleep staging is fundamental to personal health monitoring, yet many existing approaches are ill-suited for real-world applications. Traditional pipelines often rely on hand-crafted features or shallow machine learning models that struggle to generalize, while state-of-the-art deep learning methods, though accurate, are computationally heavy and impractical for resource-constrained environments. This paper introduces LightSleepX, a lightweight framework designed to deliver robust sleep analysis in resource-constrained environments. LightSleepX combines an Inception-style architecture with depthwise separable convolutions and Multi-scale Enhanced Attention for efficient multi-modal EEG/EOG feature extraction, and a Mamba encoder for rule-free long-range temporal modeling. On public benchmark datasets, LightSleepX achieves 85.9% accuracy and a 0.803 macro-F1 score on Sleep-EDF-20, and 81.8% accuracy and a 0.796 macro-F1 score on the cross-subject ISRUC-S3 dataset. With 0.049M parameters and 195.9 MFLOPs, the framework targets practical local deployment where computational cost and privacy are central constraints.
1 Introduction
Manual PSG scoring is cumbersome for continuous home tracking, while deep models can be computationally demanding and lightweight shallow methods can miss sleep-signal temporal dynamics. LightSleepX combines multi-scale EEG/EOG feature extraction with rule-free temporal modeling to target efficient local sleep staging.
1 Introduction
- Manual PSG scoring of overnight signals in 30-second epochs is cumbersome and expensive, limiting its suitability for continuous real-world home tracking.The process relies on trained experts and motivates automated sleep analysis.Source: §1 · p. 1
- Deep models can demand millions of parameters, while hand-crafted or shallow methods may fail to capture sleep's rich temporal dynamics.The passage also identifies local on-device processing as preferable in home monitoring because it avoids transmitting sensitive physiological data.Source: §1 · p. 1
- LightSleepX combines a multi-branch Inception-style extractor, depthwise separable convolutions, and Multi-scale Enhanced Attention with a Mamba encoder for rule-free long-range modeling.The approach targets efficient multi-modal EEG/EOG feature extraction and fusion.Source: §1 · p. 1
- LightSleepX is presented as a lightweight framework for local sleep staging, with 0.049M parameters and competitive benchmark performance.The introduction reports 85.9% accuracy on public benchmarks and macro-F1 scores.Source: §1 · pp. 1–2
2.1 Overall Architecture
LightSleepX processes sequences of raw EEG and EOG epochs through feature extraction, multi-modal fusion, temporal encoding, and classification. Its pipeline normalizes each epoch and produces a five-stage probability distribution.
2.1 Overall Architecture
- The architecture processes raw EEG and EOG sequences through a multi-scale feature extractor, fusion block, rule-free temporal encoder, and classification head.The four stages are presented as the model's primary components.Source: §2.1 · p. 2
- Each modality is supplied as S consecutive 30-second epochs, z-score normalized, and processed epoch-by-epoch with parallel convolutions at varying temporal resolutions.The example input has shape (B, S, 1, T), with S=10 and T=3000 samples for 100 Hz.Source: §2.1 · p. 2
- EEG and EOG feature maps are concatenated and enhanced by MEA before a bidirectional Mamba encoder and MLP produce probabilities for Wake, N1, N2, N3, and REM.MEA adaptively recalibrates channel-wise feature responses, and the encoder processes the sequence of fused feature vectors.Source: §2.1 · p. 2
2.2 Key Architectural Modules
LightSleepX processes normalized 30-second EEG/EOG epochs through multi-scale feature extraction and attention fusion, then uses bidirectional Mamba context and an MLP to classify the sequence’s center epoch.
2.2 Key Architectural Modules
- The architecture comprises a feature extractor, multi-modal fusion block, temporal encoder, and classification head.The model configuration is described as following the experimental setup.Source: §2.2 · p. 2
2.2.1 Datasets and Preprocessing
- Sleep-EDF-20, Sleep-EDF-78, and ISRUC-S3 are evaluated with standardized preprocessing and ten-epoch, five-minute input sequences.Preprocessing segments signals into 30-second epochs, merges N3 and N4, and z-score normalizes each epoch; ISRUC-S3 uses subject-based 10-fold splits.Source: §2.2.1 · p. 2
2.2.2 Multi-Branch Inception-Style Feature Extractor
- Parallel Inception-style branches capture sleep patterns at different temporal scales, while depthwise separable convolutions limit parameter cost.The implementation uses kernel sizes 3, 5, and 7 and produces 128 channels per modality.Source: §2.2.2 · pp. 2–3
2.2.3 Multi-Modal Fusion and Enhancement
- EEG and EOG features are concatenated and passed through MEA, whose parallel small- and large-kernel branches adaptively reweight features.The attention can emphasize a modality or specific features for a given epoch.Source: §2.2.3 · p. 3
2.2.4 Rule-Free Temporal Modeling with Mamba
- A bidirectional Mamba encoder models temporal dependencies using input-dependent SSM parameters and combines forward and backward context.Its selection mechanism derives parameters from the input sequence, allowing selective focus at each time step.Source: §2.2.4 · pp. 3–4
2.2.5 Classification Head
- The MLP maps each context-aware epoch feature to logits for Wake, N1, N2, N3, and REM, classifying the center epoch.The classification head is designed to avoid becoming a computational bottleneck.Source: Figure 1 · p. 3, §2.2.5 · p. 4
2.3 Experimental Setup
The setup combines Adam optimization and Dynamic Focal Loss with early stopping, then evaluates accuracy and macro F1 using subject-based cross-validation.
2.3 Experimental Setup
- Training uses batches of 128 and Adam with a 0.001 learning rate, 0.001 weight decay, and amsgrad enabled.Training runs for up to 100 epochs with validation-based early stopping.Source: §2.3 · p. 4
- Dynamic Focal Loss down-weights well-classified examples to focus training on harder samples amid sleep-data class imbalance.The focusing parameter γ is set to 2.0 and supports learning from minority classes such as N1 and N3.Source: §2.3 · p. 4
- Performance is assessed using accuracy and macro F1 across the five sleep stages, with macro F1 balancing class-wise scores.The evaluation uses subject-based K-fold splits to prevent data leakage between training and testing.Source: §2.3 · p. 4
3 Results and Discussion
LightSleepX achieves leading sleep-staging results across Sleep-EDF-20, Sleep-EDF-78, and cross-subject ISRUC-S3. Its strong performance includes N1 classification, while its parameter count is substantially smaller than several comparison models.
3.1 Comparison with State-of-the-Art Methods
- 85.9% ACC, 80.7% Macro-F1, and 0.81 κ lead all metrics on Sleep-EDF-20, while 78.7% is the highest Macro-F1 on Sleep-EDF-78.The tables compare LightSleepX with state-of-the-art methods on both datasets.Source: Table 1 · p. 4, §3.1 · p. 4, Table 2 · p. 5
3.2 Performance Analysis
- 81.80% accuracy and 0.796 Macro-F1 on cross-subject ISRUC-S3 exceed previous methods by over 5 percentage points in accuracy.The comparison covers 10-fold ISRUC-S3.Source: Table 3 · p. 5, §3.2 · p. 5
- 63.0% is LightSleepX’s state-of-the-art N1 F1-score on Sleep-EDF-78.Table 4 compares N1 F1 scores against AttnSleep, SeqSleepNet, TinySleepNet, and U-Time.Source: Table 4 · p. 5, §3.2 · p. 5
- 0.049M parameters make LightSleepX approximately 3.3 times smaller than SeqSleepNet and over 22 times smaller than U-Time and TinySleepNet.The paper frames this performance and footprint as a benchmark for the accuracy–computational-cost trade-off.Source: §3.2 · p. 5
4 Conclusion
LightSleepX combines a multi-branch Inception-style feature extractor with bidirectional Mamba for efficient sleep staging. Ablations report performance declines when proposed modules are removed or replaced.
4 Conclusion
- LightSleepX combines a Multi-Branch Inception-style feature extractor with bidirectional Mamba, achieving state-of-the-art performance with 0.049M parameters and reported strength on N1.The conclusion describes the framework as practical and privacy-preserving for local sleep monitoring.Source: §4 · p. 5
- Future work will add physiological signals such as EMG or ECG, pursue edge-hardware optimizations, and investigate semisupervised and domain-adaptation methods.The stated aim includes enhancing clinical relevance and improving generalization across diverse populations.Source: §4 · p. 5
4.1 Ablation Study
- Replacing the Inception-style block with a standard CNN decreases accuracy by 5.4%, while removing MEA declines both ACC and Macro-F1.Table 5 summarizes ablations on Sleep-EDF20, with each row removing or substituting a key module.Source: §4.1 · p. 5, Table 5 · p. 5
- Figure 2’s normalized confusion matrices show per-stage correct-classification percentages on the diagonal and misclassifications off diagonal across three datasets.The datasets are Sleep-EDF-20, Sleep-EDF-78, and ISRUC-S3.Source: Figure 2 · p. 6
- Replacing Mamba with Bi-LSTM degrades performance, and the ablation study concludes that each proposed module contributes to overall performance and efficiency.The ablation study was conducted on Sleep-EDF20.Source: §4.1 · pp. 5–6