U-Sleep: Automatic Clinical Sleep Staging

Over 15 percent of the Danish population suffers from sleep disturbances and disorders, such as sleep apnea and narcolepsy. These conditions are diagnosed using polysomnography (PSG). PSG is a time-consuming and mostly manual process, resulting in only about 4,000 PSGs being performed in Denmark each year. This is less than the current clinical need – and demand is expected to rise.

The Inspiration Behind the Innovation

PSGs measure physiological signals, which must be manually scored according to American Academy of Sleep (AASM) guidelines. This requires each 30-second window of sleep to be individually scored. A single PSG can contain up to 2,000 of such windows – amounting to up to three tedious hours of scoring per patient.

The Innovation 

U-Sleep is a highly accurate machine-learning model that can automatically detect the sleep stages in PSG data. The system’s deep neural network has been trained on a large and diverse dataset of more than 15,000 PSG studies from the United States and Europe.

Upon implementation, U-Sleep could reduce the time spent on assessing PSGs by assigning initial scores to new recordings. An expert could then spot-check the scores for accuracy (e.g., by validating 10 percent of the scorings). This could reduce the scoring time by 83–89 percent, necessitating an average of just 20 minutes per patient instead of several hours. When evaluated against consensus scores from five human experts on PSGs not used to train the model, U-Sleep was proven to be statistically non-inferior. Third-party evaluations of U-Sleep’s scoring accuracy generally found it to be superior to the evaluators’ own tools.

U-Sleep stands to reduce waiting times for PSG while improving scoring accuracy and freeing up clinicians to focus on other more patient-centric tasks. It also shows promise for adaptability for home use, further expanding the number of patients who could be screened for sleep-related conditions in the future.

The Team

Poul Jennum: Professor, Chief Physician, DMSc; Rigshospitalet

Mathias Perslev: PhD; Department of Computer Science, University of Copenhagen

Christian Igel: Professor; Department of Computer Science, University of Copenhagen

Dr. Habil: Professor; Department of Computer Science, University of Copenhagen

Miki Nikolic: PhD; Danish Center for Sleep Medicine, Rigshospitalet