DeepSleepNet-Lite: A Simplified Automatic Sleep Stage Scoring Model With Uncertainty Estimates.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Deep learning models dominate automatic sleep scoring due to high performance and direct feature learning from raw signals.
• Existing deep learning architectures are computationally intensive, process long data sequences (up to 12 minutes), and rarely offer model uncertainty estimation.

Purpose of the Study:

• To introduce DeepSleepNet-Lite, a simplified, lightweight deep learning architecture for automatic sleep scoring.
• To leverage Monte Carlo dropout for enhanced performance and uncertainty detection in sleep scoring.
• To enable real-time sleep analysis applications through a more efficient model.

Main Methods:

• Developed DeepSleepNet-Lite, a streamlined deep learning architecture processing 90-second electroencephalogram (EEG) input sequences.
• Applied Monte Carlo dropout technique for the first time in sleep scoring to estimate model uncertainty.
• Evaluated the model on the single-channel EEG Fpz-Cz from the open-source Sleep-EDF expanded database.

Main Results:

• DeepSleepNet-Lite achieved performance comparable to state-of-the-art models in accuracy, macro F1-score, and Cohen's kappa on Sleep-EDF v1 and v2.
• Monte Carlo dropout effectively estimated prediction uncertainty.
• Rejecting uncertain predictions improved model performance on both Sleep-EDF database versions.

Conclusions:

• DeepSleepNet-Lite provides a computationally efficient alternative for automatic sleep scoring.
• The integration of Monte Carlo dropout enhances model reliability by quantifying uncertainty.
• This lightweight approach facilitates real-time sleep analysis, paving the way for broader clinical and research applications.