MRASleepNet

MRASleepNet classifies sleep stages from single-channel electroencephalography (EEG) signals using integrated deep neural network modules to extract time- and frequency-domain features and model temporal relationships for clinical and research analysis.

Key Features:

• Feature Extraction Module (FE): Extracts time- and frequency-domain features from EEG signals for downstream classification.
• Multi-Resolution Attention Module (MRA): Applies attention mechanisms across multiple resolutions to refine relevant signal characteristics and reduce noise.
• Gated Multilayer Perceptron Module (gMLP): Models temporal relationships between extracted features to capture EEG dynamics over time.
• Direct Pathway for Statistical Features: Computes statistical features in parallel with deep-learning features to provide a hybrid analysis.
• Data Enhancement: Normalizes EEG signals, segments them into 30-second intervals, and incorporates contextual information from adjacent segments to produce 40-second input segments.
• Evaluation Results: Reported performance on SleepEDF-20: accuracy 84.5%, MF1 0.789, Kappa 0.786; SleepEDF-78: accuracy 81.4%, MF1 0.754, Kappa 0.743; CAP: accuracy 74.3%, MF1 0.656, Kappa 0.652.

Scientific Applications:

• Clinical Diagnostics: Automates sleep stage classification to support diagnosis and assessment of sleep disorders.
• Research Studies: Enables analysis of large EEG datasets for investigations of sleep patterns and their implications for health and disease.

Methodology:

Preprocessed EEG data are processed through the FE, MRA, and gMLP modules along with a direct statistical pathway to extract and refine features, establish multi-resolution attention, and capture temporal relationships; the network is evaluated on SleepEDF and CAP databases using accuracy, Kappa, and macro-F1 (MF1) metrics.