Unsupervised clustering of extensive physiological features substantiates five-stage sleep staging paradigm

Traditional sleep staging, guided by the American Academy of Sleep Medicine (AASM) scoring manual, categorizes sleep into five discrete stages based on visual analysis of electrophysiological signals by human expert. However, the rationale for the staging number remains underexplored, and sleep scoring results show low inter-rater agreement, due to such possible factors as subjective judgment, expertise variability among human experts, and limited number of signal features in the AASM manual. To address these limitations, we developed an unsupervised clustering framework incorporating a large set of features from electroencephalogram, electrooculogram, and electromyogram signals, including but not limited to the AASM visual features, and performing sleep staging without relying on pre-defined scoring rules. This data-driven approach shows that the sleep data can be optimally partitioned into five clusters, which correspond well to the five sleep stages defined in the AASM scoring manual. Importantly, the algorithm recognizes over 80% of AASM visual features, and additionally uncovers many features not mentioned in the AASM scoring manual. Detailed analysis into epochs inconsistently scored by the algorithm and by the human expert shows that the algorithm provides more interpretable results. The present study offers well-grounded evidence supporting that sleep should be partitioned into five stages. The findings also suggest that more features in the sleep data should be utilized in addition to those included in the AASM scoring manual for more accurate sleep scoring. Statement of Significance While the American Academy of Sleep Medicine (AASM) scoring manual services the gold standard for sleep staging, the neurophysiological basis for five rather than other number of sleep stages and the sufficiency of visual features for sleep staging remain underexplored. This study introduces an unsupervised clustering framework to explore the natural clustering in the polysomnography data. Five clusters are found to optimally classify the data, and they well correspond to the five sleep stages. The algorithm not only covers most of the AASM visual features, but also reveals critical features not visually apparent. Crucially, the extensive physiological features-based algorithm offers more interpretable staging. This study provides well-grounded evidence to support AASM's five sleep stages and highlights the necessity of expanding features for accurate sleep staging.