Sleep spindle detection using deep learning: A validation study based on crowdsourcing.

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Medicine

Background:

• Sleep spindles are key EEG oscillations in non-rapid eye movement sleep stage 2.
• Automated sleep spindle detection systems are crucial for sleep analysis.
• Deep belief networks (DBNs) show promise but are novel for sleep spindle detection.

Purpose of the Study:

• To develop and evaluate a deep belief network (DBN) for automated sleep spindle detection.
• To compare DBN performance against other classifiers using crowdsourced labeled EEG data.
• To assess feature extraction methods for DBN-based sleep spindle detection.

Main Methods:

• Generated three labeled datasets using crowdsourcing as an alternative to the gold standard.
• Trained and evaluated a deep belief network (DBN) and three other classifiers.
• Compared two power spectrum density-based feature extraction methods using DBN.

Main Results:

• The DBN achieved a high F1-score of 92.78% in classifying sleep spindle samples.
• DBN performance was comparable to expert group consensus when applied to raw EEG.
• Feature extraction methods were compared using the DBN on the same dataset.

Conclusions:

• Deep belief networks offer a promising and accurate method for automated sleep spindle detection.
• Crowdsourcing can be effectively utilized for generating labeled datasets for sleep studies.
• The developed DBN system demonstrates comparable performance to human experts in identifying sleep spindles.