Non-Linear Feature Extraction from High-Density EEG for Sleep Microstates Analysis

Conventional sleep staging, based on the AASM 30-second scoring, pro- vides a discrete representation of a process that is intrinsically continu- ous and spatially distributed. This coarse approach limits the detection of rapid transitions and local cortical variability. The present work proposes a data-driven framework for modeling sleep as a continuous and dynamic phenomenon, integrating non-linear EEG features and unsupervised prob- abilistic modeling. High-density overnight polysomnographic recordings from 29 healthy adults (ANPHY-Sleep dataset) were analyzed, including 83- channel EEG, EOG, EMG, and ECG signals. After standard pre-processing (filtering, referencing, resampling, segmentation at 30 [s] and 5 [s]), linear spectral features were complemented with the Higuchi Fractal Dimension (HFD), a non-linear descriptor capturing the temporal complexity of EEG signals. Dimensionality reduction through PCA and ICA provided compact and interpretable representations of neural activity, subsequently modeled using Gaussian Hidden Markov Models (GHMMs) at both global and re- gional levels. The unsupervised GHMM approach allowed inference of latent sleep states without manual labeling. At 30-second resolution, the model identified four states consistent with canonical AASM stages (Wake, N2, N3, REM). At higher temporal resolution (5 [s]), the optimal configura- tion expanded to eight microstates, revealing rapid within-stage transitions and transient dynamics. The inclusion of HFD increased the model’s sen- sitivity to subtle changes, improving discrimination between Wake and N2 sub-states. Model validation through leave-one-out cross-validation and comparison with the U-Sleep deep learning model demonstrated high con- sistency ( ≈ 0.6–0.8 for main stages). The regional extension, grouping electrodes into eleven cortical areas, revealed consistent temporal dimen- sionality ( = 4) but distinct spatial organization: frontal regions showed wake-like activity, centro-parietal areas were dominated by N2 dynamics, and posterior regions reflected deep-sleep patterns, consistent with the con- cept of local sleep. Overall, integrating non-linear EEG complexity measures such as HFD within unsupervised GHMM modeling provides a refined, multiscale representation of sleep microstates. This framework bridges tra- ditional discrete scoring with the continuous nature of neural dynamics, offering new insights into the spatiotemporal organization of human sleep.