SpindleU-Net: An Adaptive U-Net Framework for Sleep Spindle Detection in Single-Channel EEG

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Sleep spindles in electroencephalogram (EEG) are key biomarkers for cognitive abilities and disorders.
• Manual detection by experts is time-consuming, necessitating automated methods.
• Existing automated methods show room for performance improvement.

Purpose of the Study:

• To develop an automated, high-performance method for detecting sleep spindles in EEG.
• To leverage deep learning and image segmentation techniques for spindle detection.
• To improve upon current state-of-the-art automated sleep spindle detection.

Main Methods:

• Proposed SpindleU-Net, a U-Net framework with an attention module for point-wise spindle detection.
• Input EEG sequences of arbitrary length are mapped to dense spindle/non-spindle labels.
• An attention module focuses on salient spindle regions, and a task-specific loss function addresses class imbalance.

Main Results:

• SpindleU-Net outperformed state-of-the-art methods on the MASS and DREAMS EEG datasets.
• Achieved average F1 scores of 0.854 and 0.803 on the MASS dataset (vs. two experts).
• Obtained an average F1 score of 0.739 on the DREAMS dataset, demonstrating good cross-dataset generalization.

Conclusions:

• SpindleU-Net offers a significant advancement in automated sleep spindle detection.
• The method shows strong performance and generalization capabilities across different datasets.
• This deep learning approach enhances clinical research by providing efficient and accurate biomarker analysis.