Deep clustering of polysomnography data to characterize sleep structure in healthy sleep and non-rapid eye movement parasomnias

• Published in: Journal of neuroscience methods Vol. 423; p. 110516
• Main Authors: Huijben, I.A.M., van Sloun, R.J.G., Pijpers, A., Overeem, S., van Gilst, M.M.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.11.2025
• Subjects: Clustering; Dimensionality reduction; Disorders of arousal; EEG analysis; NREM parasomnia; Polysomnography; Sleep; Dimensionality reduction; NREM parasomnia; Sleep; EEG analysis; Disorders of arousal; Clustering; Polysomnography
• ISSN: 0165-0270; 1872-678X; 1872-678X
• DOI: 10.1016/j.jneumeth.2025.110516

The clinical standard to interpret polysomnography (PSG) data is to categorize sleep in five stages, which omits information. SOM-CPC is an unsupervised method that extracts features through contrastive predictive coding (CPC), and visualizes them in two dimensions using a self-organizing map (SOM). We propose various visualizations and analyses for pattern recognition in PSG data through SOM-CPC. We used SOM-CPC to learn a representation of 30-s multi-channel epochs from two datasets of healthy sleepers (n=52 and n=22 in the test sets). SOM-CPC was, additionally, used to further characterize awakenings from slow wave sleep (SWS) in non-rapid eye movement (NREM) parasomnias. For the latter, SOM-CPC was trained on 5-s single-channel EEG windows of non-rapid eye movement parasomnias and matched healthy controls (test set: n=67). SOM-CPC organized epochs of healthy sleepers such that it separated sleep stages, and also encoded age of the subjects and time in the night. Parasomnia episodes, compared to non-behavioral SWS awakenings, exhibited higher SWS-specificity prior to transition to wakefulness, higher Wake-specificity post-transition, and longer durations. The learned representations were compared against gold-standard sleep stage labels and variables known to impact sleep structure. SOM-CPC seems a useful model for pattern discovery in PSG data, as it enables observation of state changes that are more intricate than full sleep stage transitions. It, moreover, provided further evidence for signal level differences in the EEG between SWS awakenings with and without parasomnia episodes. •Representation learning of polysomnography data enables pattern discovery.•The unsupervised SOM-CPC model visualizes polysomnography data in two dimensions.•In healthy sleepers, the 2D representation coincided well with sleep stages.•The 2D representation enables comparison of different variables to find patterns.•In NREM parasomnias, SOM-CPC further characterized the EEG of behavioral episodes.