Applicability of LSTM and Double Decoders on U-Net models for Sleep and Arousal Scoring Using In-Home EEG Signals

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Machine learning is increasingly used for sleep analysis via polysomnography (PSG).
• In-home electroencephalography (EEG) analysis for sleep and arousal scoring faces challenges.
• Arousal detection is more complex than sleep stage classification, requiring specialized methods.

Purpose of the Study:

• To explore multitask learning using U-Net architectures for joint sleep and arousal scoring in in-home EEG data.
• To evaluate architectural innovations, including bidirectional LSTM layers and branching double decoders, within U-Net models.
• To assess the performance and robustness of these models across diverse datasets.

Main Methods:

• Utilized U-Net architectures with integrated bidirectional LSTM layers.
• Implemented a branching double decoder for task-specific outputs.
• Trained and evaluated models on four datasets, including those with sleep apnea and low-noise EEG.

Main Results:

• Specific U-Net configurations (SAAS U-Net DD-L, D-L) excelled in arousal scoring.
• Other configurations (SAAS U-Net L-DD, L-D) showed superior performance in sleep scoring.
• Fine-tuning strategies further enhanced sleep scoring accuracy, demonstrating model robustness.

Conclusions:

• Multitask learning and U-Net architectural enhancements are valuable for in-home EEG sleep diagnostics.
• Proposed models show practical potential for improving sleep and arousal scoring accuracy.
• These advancements can lead to more reliable portable sleep disorder diagnostics.