Few-shot EEG sleep staging based on transductive prototype optimization network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for monitoring brain activity and diagnosing sleep disorders.
• Manual sleep staging from EEG signals is time-consuming and requires expert labor.
• Current deep learning methods struggle with limited data for specific sleep stages.

Purpose of the Study:

• To develop an improved EEG sleep staging method using few-shot learning.
• To introduce the transductive prototype optimization network (TPON) for enhanced sleep stage classification.
• To address the challenge of limited labeled data in EEG sleep staging.

Main Methods:

• Proposed a novel few-shot learning approach named transductive prototype optimization network (TPON).
• Utilized meta-learning algorithms to enable generalization to new classes with few examples.
• Optimized prototype distribution using support sets and high-confidence unlabeled samples for improved authenticity.

Main Results:

• The TPON method demonstrated superior performance in EEG sleep staging compared to advanced algorithms on the SleepEDF-2013 dataset.
• Achieved high accuracy in sleep stage classification, validating the effectiveness of the few-shot learning approach.
• Experimentally confirmed the feasibility of cross-channel recognition, suggesting shared sleep EEG features across channels.

Conclusions:

• The TPON method offers an effective solution for accurate EEG sleep staging, especially in data-limited scenarios.
• The findings highlight the potential of meta-learning and prototype optimization for medical signal analysis.
• Future research can explore universal features across EEG channels for further improvements in sleep analysis.