ECG-SleepNet: Deep Learning-Based Comprehensive Sleep Stage Classification Using ECG Signals

TL;DR

This paper tackles five-class sleep stage classification from a single-channel ECG, addressing the invasive and complex nature of PSG-based approaches. It introduces ECG-SleepNet, a three-stage framework that combines a Feature Imitating Network (FIN) for robust feature initialization, a dedicated N1 detector based on time-frequency SST representations, and a final Kolmogorov-Arnold Network (KAN)–based fusion to classify Wake, N1, N2, N3, and REM. Key contributions include the ECG-only modality, improved N1 detection through stage-specific pretraining, and the use of data augmentation (notably SMOTE) to balance classes, achieving an overall accuracy of $80.79\%$ and a Cohen's kappa of $0.73$ on the MIT-BIH dataset, with Wake/N1/N2/N3/REM per-class accuracies approaching practical utility for wearable sleep monitoring. The approach demonstrates real-time feasibility (latency ≈ $97.77$ ms per 30-second ECG) and provides a foundation for accessible, noninvasive sleep health assessment in home or wearable settings. The results highlight the importance of weight initialization, multi-stage specialization, and learnable-activation networks in capturing the nuanced patterns of sleep-stage transitions from ECG alone.

Abstract

Accurate sleep stage classification is essential for understanding sleep disorders and improving overall health. This study proposes a novel three-stage approach for sleep stage classification using ECG signals, offering a more accessible alternative to traditional methods that often rely on complex modalities like EEG. In Stages 1 and 2, we initialize the weights of two networks, which are then integrated in Stage 3 for comprehensive classification. In the first phase, we estimate key features using Feature Imitating Networks (FINs) to achieve higher accuracy and faster convergence. The second phase focuses on identifying the N1 sleep stage through the time-frequency representation of ECG signals. Finally, the third phase integrates models from the previous stages and employs a Kolmogorov-Arnold Network (KAN) to classify five distinct sleep stages. Additionally, data augmentation techniques, particularly SMOTE, are used in enhancing classification capabilities for underrepresented stages like N1. Our results demonstrate significant improvements in the classification performance, with an overall accuracy of 80.79% an overall kappa of 0.73. The model achieves specific accuracies of 86.70% for Wake, 60.36% for N1, 83.89% for N2, 84.85% for N3, and 87.16% for REM. This study emphasizes the importance of weight initialization and data augmentation in optimizing sleep stage classification with ECG signals.

Figures (5)

• Figure 1: Overview of the three-stage sleep stage classification model. Stage 1 involves a FIN, utilizing an LNN architecture, trained on synthetic signals to estimate kurtosis and skewness. Stage 2 employs an architecture inspired by ResUNet++ to perform binary classification of N1 versus non-N1 stages using SST-based TFRs. In Stage 3, the pretrained FIN and the model used in Stage 2 are integrated using KANs for the final classification into Wake, N1, N2, N3, and REM stages.
• Figure 2: Detailed components of the ASPP, ResNet, SE, and STEM blocks used in the the encoder block.
• Figure 3: Predicted vs. True Values for Skewness (Top) and Kurtosis (Bottom).
• Figure 4: Validation Accuracy with FIN Initialization vs. without FIN Initialization.
• Figure 5: Confusion matrix of the sleep stage classification model on the test dataset. Each row represents the true label, and each column represents the predicted label. The diagonal values indicate correct classifications, while off-diagonal values represent misclassifications between sleep stages.