Sleep is a dynamic physiological process whose spatio-temporal complexity far exceeds the dynamics that can be captured by traditional clinical staging, which is based on assigning a single global state to 30-second epochs according to the American Academy of Sleep Medicine guidelines. This thesis work aims to validate the applicability of the Gaussian-Linear Hidden Markov Model framework with Time-Delay Embedding (TDE-HMM) to full-night polysomnographic recordings, lasting 7-8 hours and acquired using an 83-channel electroencephalographic system according to the extended 10-10 international system. The dataset used is ANPHY-Sleep, an open-access database containing overnight recordings of 29 healthy subjects, provided with certified manual staging. The methodological pipeline involves a rigorous pre-processing phase in the MATLAB environment (EEGLAB), including zero-phase high-pass and low-pass filtering (0.3-45 Hz) to eliminate slow drifts and muscle artifacts, downsampling to 125 Hz, and artifact removal via Independent Component Analysis integrated with visual inspection. The pre-processed signal is subsequently analyzed in Python using the glhmm package, applying the TDE-HMM model with a 100 ms embedding window and PCA dimensionality reduction to 20 principal components. The choice of TDE is motivated by the stringent computational constraints inherent in the high density of the EEG montage. The framework was initially validated on a single subject, systematically exploring the hidden state space ($K \in [2, 100]$) and confirming the profound neurophysiological coherence of the networks extracted for $K = 5$ and $K = 10$ through the analysis of power spectra and band topographies. Subsequently, the model was trained on 10 temporally concatenated subjects, obtaining a robust and generalizable population description. The results demonstrate that the TDE-HMM identifies, in a completely unsupervised manner, microstates with widely replicable cross-subject topographic signatures. Among these, a fronto-central Delta-dominant configuration (consistent with slow-wave sleep) and a state with an isolated peak in the centro-occipital Alpha band (typical of relaxed wakefulness) emerge for their maximum stability. The inter-individual variance is selectively concentrated in the frequency bands expected from the clinical literature, confirming the extraction of real neurobiological dynamics and not mere statistical artifacts. The main contribution of this study is having demonstrated the effective scalability of the TDE-HMM framework to long-duration, high-density EEG datasets. The adopted approach yields a description of the sleep microstructure at an extremely high temporal resolution and defines the analytical infrastructure preparatory for future investigations into state-specific transition trajectories and the topographic quantification of the local sleep phenomenon.
Il sonno è un processo fisiologico dinamico la cui complessità spazio-temporale eccede ampia- mente le dinamiche catturabili dalla stadiazione clinica tradizionale, basata sull’assegnazione di un unico stato globale a epoche di 30 secondi secondo le linee guida American Academy of Sleep Medicine. Il presente lavoro di tesi si propone di validare l’applicabilità del framework Gaussian-Linear Hidden Markov Model con Time-Delay Embedding (TDE-HMM) a registra- zioni polisonnografiche notturne integrali, della durata di 7-8 ore e acquisite con un sistema elettroencefalografico a 83 canali secondo il sistema internazionale 10-10 esteso. Il dataset utilizzato è ANPHY-Sleep, un database open access contenente registrazioni notturne di 29 sog- getti sani, provviste di stadiazione manuale certificata. La pipeline metodologica prevede una rigorosa fase di pre-elaborazione in ambiente MATLAB (EEGLAB) comprendente filtraggio passa-alto e passa-basso (0.3-45 Hz) a fase zero per l’eliminazione delle derive lente e degli artefatti muscolari, sottocampionamento a 125 Hz e rimozione degli artefatti tramite Analisi delle Componenti Indipendenti integrata da ispezione visiva. Il segnale pre-elaborato è succes- sivamente analizzato in Python tramite il pacchetto glhmm, applicando il modello TDE-HMM con finestra di embedding di 100 ms e riduzione dimensionale PCA a 20 componenti prinici- pali. La scelta del TDE è motivata dai stringenti vincoli computazionali intrinseci all’elevata densità del montaggio EEG. Il framework è stato inizialmente validato su un singolo soggetto, esplorando sistematicamente lo spazio degli stati nascosti ( ∈ [2, 100]) e confermando la pro- fonda coerenza neurofisiologica delle reti estratte per = 5 e = 10 attraverso l’analisi degli spettri di potenza e delle topografie di banda. Successivamente, il modello è stato addestrato su 10 soggetti concatenati temporalmente, ottenendo una descrizione di popolazione robusta e generalizzabile. I risultati dimostrano che il TDE-HMM identifica, in modalità completamente non supervisionata, microstati con firme topografiche ampiamente replicabili cross-subject. Tra questi, emergono per massima stabilità, una configurazione a dominanza Delta fronto-centrale (coerente con il sonno ad onde lente) e uno stato con isolato picco in banda Alpha centro- occipitale (tipico della veglia rilassata). La varianza interindividuale risulta selettivamente concentrata nelle bande di frequenza attese dalla letteratura clinica, confermando l’estrazione di dinamiche neurobiologiche reali e non di meri artefatti statistici. Il contributo principale di questo studio consiste nell’aver dimostrato l’effettiva scalabilità del framework TDE-HMM a dataset EEG ad alta densità di lunga durata. L’approccio adottato restituisce una descrizione della microstruttura del sonno ad altissima risoluzione temporale e definisce l’infrastruttura analitica propedeutica per future indagini sulle traiettorie di transizione stato-specifiche e per la quantificazione topografica del fenomeno del sonno locale.
Analisi data-driven della microstruttura del sonno mediante Time-Delay Embedded Hidden Markov Model applicato a EEG notturno ad alta densità
MATTII, ELENA
2025/2026
Abstract
Sleep is a dynamic physiological process whose spatio-temporal complexity far exceeds the dynamics that can be captured by traditional clinical staging, which is based on assigning a single global state to 30-second epochs according to the American Academy of Sleep Medicine guidelines. This thesis work aims to validate the applicability of the Gaussian-Linear Hidden Markov Model framework with Time-Delay Embedding (TDE-HMM) to full-night polysomnographic recordings, lasting 7-8 hours and acquired using an 83-channel electroencephalographic system according to the extended 10-10 international system. The dataset used is ANPHY-Sleep, an open-access database containing overnight recordings of 29 healthy subjects, provided with certified manual staging. The methodological pipeline involves a rigorous pre-processing phase in the MATLAB environment (EEGLAB), including zero-phase high-pass and low-pass filtering (0.3-45 Hz) to eliminate slow drifts and muscle artifacts, downsampling to 125 Hz, and artifact removal via Independent Component Analysis integrated with visual inspection. The pre-processed signal is subsequently analyzed in Python using the glhmm package, applying the TDE-HMM model with a 100 ms embedding window and PCA dimensionality reduction to 20 principal components. The choice of TDE is motivated by the stringent computational constraints inherent in the high density of the EEG montage. The framework was initially validated on a single subject, systematically exploring the hidden state space ($K \in [2, 100]$) and confirming the profound neurophysiological coherence of the networks extracted for $K = 5$ and $K = 10$ through the analysis of power spectra and band topographies. Subsequently, the model was trained on 10 temporally concatenated subjects, obtaining a robust and generalizable population description. The results demonstrate that the TDE-HMM identifies, in a completely unsupervised manner, microstates with widely replicable cross-subject topographic signatures. Among these, a fronto-central Delta-dominant configuration (consistent with slow-wave sleep) and a state with an isolated peak in the centro-occipital Alpha band (typical of relaxed wakefulness) emerge for their maximum stability. The inter-individual variance is selectively concentrated in the frequency bands expected from the clinical literature, confirming the extraction of real neurobiological dynamics and not mere statistical artifacts. The main contribution of this study is having demonstrated the effective scalability of the TDE-HMM framework to long-duration, high-density EEG datasets. The adopted approach yields a description of the sleep microstructure at an extremely high temporal resolution and defines the analytical infrastructure preparatory for future investigations into state-specific transition trajectories and the topographic quantification of the local sleep phenomenon.| File | Dimensione | Formato | |
|---|---|---|---|
|
Mattii_Elena.pdf
embargo fino al 19/04/2029
Dimensione
25 MB
Formato
Adobe PDF
|
25 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/107650