Sleep stage classification from ECG using machine learning: Evaluating the impact of signal duration

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Medical Signal Processing

Background:

• Accurate sleep stage classification is vital for diagnosing sleep disorders like insomnia and sleep apnea.
• Current manual scoring methods are time-consuming and limit scalability, especially in resource-limited settings.
• Automating sleep stage classification using machine learning can enhance diagnostic efficiency and reduce healthcare burdens.

Purpose of the Study:

• To develop and evaluate a simplified sleep stage classification system using only electrocardiogram (ECG) signals.
• To assess the feasibility of using heart rate variability (HRV) and Poincaré plot descriptors for sleep stage prediction.
• To demonstrate a cost-effective and scalable solution for sleep monitoring, particularly in low-resource environments.

Main Methods:

• Extracted heart rate variability (HRV) and Poincaré plot features from ECG signals.
• Trained machine learning models (Neural Networks, KNN, XGBoost, Random Forest) for five-stage sleep classification.
• Validated the approach on two public datasets (Haaglanden Medisch Centrum and MIT-BIH Polysomnographic Databases).

Main Results:

• The highest classification accuracy achieved was 67% using long-duration ECG recordings.
• Models trained on longer ECG segments outperformed those trained on shorter segments by 12%.
• The study highlights the significant impact of signal duration on classification performance.

Conclusions:

• Sleep stage classification is feasible using solely ECG signals, offering a simpler and more accessible alternative to multi-signal methods.
• ECG-only systems demonstrate potential for portable, low-cost, and scalable sleep monitoring solutions.
• The findings support the development of more accessible sleep disorder detection tools for low-resource settings.