GTransU-CAP: Automatic labeling for cyclic alternating patterns in sleep EEG using gated transformer-based U-Net...

Area of Science:

• Sleep Medicine
• Computational Neuroscience
• Biomedical Engineering

Background:

• Cyclic alternating pattern (CAP) sequences, characterized by alternating activation (A-phases) and background phases, are crucial bio-markers of sleep instability.
• CAP A-phase subtypes are linked to sleep pathologies, making their accurate detection clinically significant.
• Manual labeling of CAP A-phases by sleep experts is time-consuming and burdensome.

Purpose of the Study:

• To develop an automated method for detecting CAP A-phases and their subtypes.
• To leverage the unique characteristics of CAP A-phases for improved detection performance.
• To reduce the manual workload in sleep analysis.

Main Methods:

• Proposed an automated detection method utilizing a Transformer-based U-Net framework.
• Integrated local feature extraction (U-Net) with global dependency modeling (Transformer) to capture long-span A-phase characteristics.
• Employed a curriculum-learning strategy to enhance model training and performance.

Main Results:

• Achieved average F1 scores of 67.78% for A-phase detection in healthy subjects and 72.16% in patients with nocturnal frontal lobe epilepsy.
• Obtained an average macro F1-score of 59.5% for multi-class CAP A-phase subtype classification.
• Demonstrated superior performance compared to existing state-of-the-art methods in both A-phase detection and subtype classification.

Conclusions:

• The proposed Transformer-based U-Net method effectively automates the detection and classification of CAP A-phases and their subtypes.
• The approach successfully exploits CAP characteristics, offering improved accuracy over previous methods.
• This automated system holds significant potential for clinical application in sleep disorder diagnosis and research.