Beyond K-complex binary scoring during sleep: probabilistic classification using deep learning.

Area of Science:

• Neuroscience
• Sleep Medicine
• Biomedical Engineering

Background:

• K-complexes (KCs) are key electroencephalography (EEG) markers of sleep stage 2, reflecting sensory processing.
• KC characteristics may indicate sleep quality, aging, and various neurological deficits.
• Manual KC scoring is labor-intensive, costly, and lacks standardization, limiting research.

Purpose of the Study:

• To develop and validate a novel, automated algorithm for accurate K-complex detection in sleep EEG.
• To improve upon the performance and accessibility of existing automated KC scoring methods.

Main Methods:

• A deep neural network and Gaussian process model was employed for probabilistic KC classification (0-100% probability).
• The algorithm was trained on 500,000 synthetic KCs from manually scored sleep stage 2 data.
• Performance was evaluated on 700 independent sleep recordings from the Cleveland Family Study, using sleep stages 2 and 3.

Main Results:

• The developed algorithm achieved a high F1 score of 0.78, significantly outperforming existing methods (F1 = 0.2-0.6).
• The probabilistic output effectively captured individual and age-related variations in KC morphology and amplitude.

Conclusions:

• Automated probabilistic K-complex classification provides an effective and systematic method for KC detection.
• This tool facilitates deeper investigation into the links between KCs, sleep, and clinical outcomes like health impacts and daytime symptoms.