AttnSleep

AttnSleep classifies sleep stages from single-channel electroencephalogram (EEG) recordings using an attention-based deep learning architecture to enable accurate automatic sleep stage classification.

Key Features:

• Multi-Resolution Convolutional Neural Network (MRCNN): Employs an MRCNN to extract both low- and high-frequency features from single-channel EEG signals.
• Adaptive Feature Recalibration (AFR): Models inter-dependencies among extracted features and recalibrates feature responses to emphasize relevant signals for classification.
• Temporal Context Encoder (TCE) with Multi-Head Attention: Uses multi-head attention combined with causal convolutions to capture temporal dependencies within extracted features.
• Two-stage processing: Separates analysis into feature extraction (MRCNN + AFR) followed by temporal modeling (TCE) for classification.

Scientific Applications:

• Sleep stage classification: Provides automatic sleep staging from single-channel EEG for sleep research and clinical studies.
• Sleep disorder research: Supports investigation of sleep disorders by providing automated stage annotations.
• Sleep quality assessment: Enables measurement and assessment of sleep quality through automated staging.
• Neurological condition studies: Supports research linking sleep-stage patterns to neurological conditions.
• Benchmarking and evaluation: Has been validated across three public datasets and demonstrated improved performance over existing state-of-the-art techniques in evaluation metrics.

Methodology:

Two-stage computational pipeline: feature extraction using MRCNN combined with AFR to extract and refine features from single-channel EEG, followed by temporal analysis using a TCE employing multi-head attention and causal convolutions to model temporal dependencies for classification.