0431 The Hypno-PC: Uncovering Sleep Dynamics Through Unsupervised Learning

Abstract Introduction Traditional sleep research primarily depends on visually scored stages based on electrophysiological signals. This manual approach is time-consuming and subject to subjective biases, implicitly assuming that visually defined categories accurately represent the underlying biological processes. Recent advancements, however, have shown that complex brain activity can be effectively represented in low-dimensional spaces, offering valuable insights into the temporal organization and structure of physiological states. This raises the question of how such data-driven representations relate to and potentially refine our conventional understanding of sleep structure. Methods In this study, we developed a data-driven framework for identifying inherent brain states directly from continuous physiological signals. We applied Principal Component Analysis (PCA) to features extracted from overnight high-density EEG, EOG, EMG, and ECG recordings at 30-second and 4-second resolutions. After identifying the principal axes of variation, we employed a Gaussian Hidden Markov Model (GHMM) on the PCA-transformed data to delineate discrete states. To align the hidden states with the sleep labels, we used a minimally supervised approach—less than 0.5% of labeled data—and a cross-subject approach. Results The first principal component (PC1), termed the “Hypno-PC,” showed a strong correspondence with the manually scored hypnogram, indicating that the largest source of variance in the spectral profile aligns closely with standard sleep staging. Furthermore, the GHMM-derived states achieved an agreement with conventional sleep labels at a level comparable to the reported inter-rater agreement. These results suggest that the states identified through purely data-driven means closely mirror established concepts of sleep architecture. Preliminary results on data from individuals with epilepsy indicate that this low-dimensional representation may reflect not only physiological but also pathological processes. Conclusion By integrating PCA and GHMM, we present a reproducible, scalable, and flexible methodology that complements traditional sleep scoring methods. While demonstrated with sleep and spectral features, this approach is adaptable to other continuous physiological signals. The findings support the notion that unsupervised, data-driven methods can uncover intrinsic patterns and structures in both normal and abnormal states. This perspective promotes a more nuanced understanding of the inherent organization of physiological and pathological state dynamics. Support (if any)