The Hypno-PC: Uncovering Sleep Dynamics through Principal Component Analysis and Hidden Markov Modeling of Electrophysiological Signals

Abstract

The conventional approach to sleep analysis relies on pre-defined, visually scored stages derived from electrophysiological signals. This manual method demands substantial effort and is influenced by subjective assessments, implicitly assuming that these categories accurately reflect underlying biological processes. Recent advancements indicate that low-dimensional representations of complex brain activity can provide objective means of identifying brain states. These approaches can potentially uncover inherent patterns within sleep, offering valuable insights into its organization. In this study, we applied Principal Component Analysis (PCA) to spectral features extracted from high-density EEG, EOG, EMG, and ECG recorded overnight at both 30- and 4-second resolutions. Notably, the first principal component--the "Hypno-PC"--strongly aligns with the hypnogram at both time scales. Subsequently, we employed a Gaussian Hidden Markov Model (GHMM) to delineate discrete states in the PCA-transformed data and to quantify their temporal dynamics. Using minimal supervision (less than 0.5% of the data labeled) and a cross-subject approach, the model achieved alignment with standard sleep labels comparable to the typical inter-rater agreement. Finally, independent component analysis (ICA) was applied to the PCA space, decomposing it into an independent set of components that potentially represent distinct physiological processes. The integrated use of PCA, GHMM, and ICA provides a reproducible and scalable methodology that aligns with traditional sleep staging, while offering a more flexible and comprehensive perspective on sleep states. Our findings indicate that these data-driven, unsupervised methods effectively uncover the intrinsic dynamics of sleep, advancing automated sleep analysis and enhancing our understanding of sleep organization. Statement of SignificanceThis study introduces a data-driven framework for sleep analysis designed to objectively identify brain states using low-dimensional representations of electrophysiological signals during sleep. By integrating a suite of unsupervised learning techniques, our methodology offers an alternative to subjective manual scoring, potentially enhancing both efficiency and reproducibility. In addition to aligning with traditional sleep staging, this approach uncovers subtle sleep dynamics across multiple time scales, enabling the discovery of patterns that conventional methods might overlook. These advancements hold promise for automated sleep monitoring and the study of sleep disorders, potentially improving diagnostic accuracy and facilitating large-scale sleep research. By addressing current limitations in sleep analysis techniques, this framework lays the groundwork for more elaborate and scalable assessments.