Abstract
Sleep quality assessment is vital for diagnosing sleep-related pathologies, yet standard polysomnography (PSG) is complex and uncomfortable for home use. This study introduces a sleep stage classification system utilizing only two forehead EEG channels (FP1 and FP2). By leveraging clinical expert knowledge and a Relevance Vector Machine (RVM) classifier, the system achieves an average accuracy of 76.7%, offering a practical alternative for long-term home sleep monitoring.
Introduction
Polysomnography (PSG) remains the "gold standard" for sleep analysis, typically requiring at least 11 channels to monitor brain activity (EEG), eye movements (EOG), muscle tone (EMG), and heart rate (ECG). While highly accurate, the physical burden of these sensors and the need for manual scoring by clinicians make PSG unsuitable for routine home use. There is a significant need for a simplified, automated system that can reliably stage sleep without the interference of hair or the discomfort of extensive wiring.
Methodology
The proposed system focuses on two forehead EEG channels, FP1 and FP2. The forehead is an ideal recording site because it is hair-free, allowing for easy application via headbands or self-adhesive sensors.
Feature Engineering
Rather than relying solely on black-box deep learning, the researchers engineered features based on the American Academy of Sleep Medicine (AASM) manual scoring rules. To capture transient events like sleep spindles, which are often lost in standard 30-second Fast Fourier Transforms (FFT), the team employed Short-Time Fourier Transform (STFT). Each 30-second epoch was divided into 59 overlapping segments, allowing the system to identify specific spectral powers (Delta, Theta, Alpha, Sigma, Beta, and Gamma) that correspond to clinical sleep stage markers.
Classification via Relevance Vector Machine (RVM)
The system utilizes an RVM classifier, a Bayesian alternative to the Support Vector Machine (SVM). RVMs are known for producing "sparser" models, using fewer data points (relevance vectors) to define decision boundaries. This reduces the risk of overfitting and improves computational efficiency, which is critical for potential integration into wearable devices.
Results and Discussion
The system was validated using a Leave-One-Subject-Out (LOSO) cross-validation approach on data from 10 healthy male participants.
Performance Metrics
The RVM-based system achieved an average accuracy of 76.7% ± 4.0% and a Cohen’s Kappa of 0.68, indicating substantial agreement with human-scored PSG. Notably, the system excelled at identifying deep sleep stages, with sensitivities of 85.6% for N2 and 87.9% for N3.
Advantages of Forehead EEG
A key finding was that forehead EEG acts as a "hybrid" signal. Due to the proximity to the eyes, these channels naturally capture EOG data, which is essential for identifying the Rapid Eye Movement (REM) stage without additional sensors. Furthermore, the RVM significantly outperformed traditional SVMs, maintaining high accuracy while using a fraction of the support vectors.
Limitations
The system faced challenges common to automated sleep staging, particularly in identifying Stage N1 (23.2% sensitivity), which shares overlapping characteristics with wakefulness and REM sleep. Additionally, the small, homogeneous sample size (10 young males) means further testing is required to ensure the model's generalizability across different ages and sleep pathologies.
Conclusion
This research demonstrates that a two-channel forehead EEG system, combined with expert-knowledge-based features and RVM classification, provides a viable and convenient alternative to full PSG for home sleep monitoring. By simplifying the hardware requirements while maintaining clinical relevance, this approach paves the way for accessible, long-term sleep health management.
Summary updated