A real-time automated sleep scoring algorithm to detect refreshing sleep in conscious ventilated critically ill patients.

Objectives: Due to the noisy environment, a very large number of patients admitted to intensive care units (ICUs) suffer from sleep severe disruption. These sleep alterations have been associated with a prolonged need for assisted ventilation or even with death. Sleep scoring in the critically ill is very challenging and requires sleep experts, limiting relevant studies to a few experienced teams. In this context, an automated scoring system would be of interest for researchers. In addition, real-time scoring could be used by nurses to protect patients' sleep. We devised a sleep scoring algorithm working in real time and compared this automated scoring against visual scoring.

Methods: We analyzed retrospectively 45 polysomnographies previously recorded in non-sedated and conscious ICU patients during their weaning phase. For each patient, one EEG channel was processed, providing automated sleep scoring. We compared total sleep time obtained with visual scoring versus automated scoring. The proportion of sleep episodes correctly identified was calculated.

Results: Automated total sleep time and visual sleep time were correlated; the automatic system overestimated total sleep time. The median [25th-75th] percentage of sleep episodes lasting more than 10 min detected by algorithm was 100% [73.2 - 100.0]. Median sensitivity was 97.9% [92.5 - 99.9].

Conclusion: An automated sleep scoring system can identify nearly all long sleep episodes. Since these episodes are restorative, this real-time automated system opens the way for EEG-guided sleep protection strategies. Nurses could cluster their non-urgent care procedures, and reduce ambient noise so as to minimize patients' sleep disruptions.