A robust deep learning detector for sleep spindles and K-complexes: towards population norms
Nicolás I Tapia-Rivas1, Pablo A Estévez2,3,4, José A Cortes-Briones5,6,7
1Department of Electrical Engineering, University of Chile, Santiago, Chile.
Scientific Reports
|January 3, 2024
Summary
A new deep learning system, the Sleep EEG Event Detector (SEED), accurately identifies sleep spindles and K-complexes. This advanced tool aids large-scale sleep studies by improving the precision of sleep event detection and analysis.
Area of Science:
- Neuroscience
- Computational Biology
- Sleep Medicine
Background:
- Sleep spindles (SSs) and K-complexes (KCs) are crucial electroencephalography (EEG) patterns during sleep, linked to cognitive functions.
- Accurate detection of SSs and KCs is vital for large-scale sleep research but challenging due to signal variability and annotation inconsistencies.
Purpose of the Study:
- To introduce the Sleep EEG Event Detector (SEED), a deep learning system for robust automatic detection of SSs and KCs.
- To demonstrate SEED's superior performance compared to existing methods and its adaptability to new datasets and annotation styles.
Main Methods:
- Development of SEED, a deep learning system utilizing a novel pretraining approach with the A7 rule-based detector to reduce data requirements.
- Evaluation of SEED on the MASS2 dataset, comparing its performance against established detection approaches.
Main Results:
- SEED achieved high F1-scores of 80.5% for SS detection and 83.7% for KC detection on the MASS2 dataset.
- SEED demonstrated strong transferability, requiring minimal fine-tuning for different datasets and annotation styles.
- Analysis of 11,224 subjects showed SEED provides improved estimates of SS population statistics.
Conclusions:
- SEED is a powerful and adaptable deep learning tool for precise automatic detection of sleep EEG events.
- The system's efficiency in data annotation and improved statistical estimation make it valuable for sleep research and establishing population norms.
