A Few-Shot Learning-Based EEG and Stage Transition Sequence Generator for Improving Sleep Staging Performance.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Sleep Medicine

Background:

• Automatic sleep stage classification is crucial for diagnosing sleep disorders.
• High accuracy requires extensive labeled electroencephalogram (EEG) data, which is costly and time-consuming to acquire.
• Few-shot learning approaches are being explored to address data scarcity.

Purpose of the Study:

• To propose SleepGAN, a generative adversarial network (GAN) framework, for augmenting sleep EEG datasets.
• To generate realistic EEG epochs and sleep stage transition sequences using few-shot learning.
• To evaluate the impact of generated data on the performance of automatic sleep stage classification models.

Main Methods:

• Developed progressive Wasserstein divergence GANs for EEG epoch generation.
• Utilized a relational memory generator for sleep stage transition sequence generation.
• Employed few-shot learning principles to train GANs with limited data.
• Evaluated generated data using single-channel EEGs from the Sleep-EDF dataset.

Main Results:

• The inclusion of generated EEG epochs improved classification accuracy by approximately 1%.
• Adding both generated EEG epochs and stage transition sequences increased accuracy by 3%, reaching 83.06% from 79.40%.
• SleepGAN effectively generated realistic data and sequences, demonstrating its utility in data augmentation.

Conclusions:

• SleepGAN successfully generates realistic EEG epochs and transition sequences, even with insufficient training data.
• The proposed method enhances sleep stage classification model performance through data augmentation.
• SleepGAN offers a viable solution for improving clinical practice in sleep analysis by overcoming data limitations.