Sleep represents an essential physiological process for human health, the fragmentation of which impacts the various vital functions it performs for the body and consequently causes numerous systemic disorders. Among the main mechanisms of disruption is autonomic arousal, which involves brief and transient activations of the autonomic nervous system that are not always accompanied by EEG signals and are therefore difficult to detect. In this thesis, an automatic framework was developed for the noninvasive detection of these events, based solely on tachogram analysis (time series of RR intervals), with the future goal of application in the wearable domain via PPG-derived signals. Starting from the PhysioNet 2018 dataset, several pre-processing stages were implemented, including ECG signal filtering, R-peak detection, tachogram construction, signal quality assessment, and segmentation into 60-second windows. An element of particular innovativeness lies in the construction of the test set: while the models were trained on EEG annotations related to central arousal, the evaluation phase was conducted on an independent set manually annotated through tachographic signal alone, to detect autonomic activation events. This approach realistically simulates use scenarios without EEG, such as those related to wearable devices. Three deep learning models (CNN-LSTM, UNet1D, and Transformer) were designed and compared for binary classification of segments, evaluated with optimized decision thresholds, and on either using only high-quality segments or using all segments indiscriminately. Finally, a post-processing pipeline based on morphological filters allowed a significant reduction in false positives (up to 78%), with increased F1-score and accuracy, especially for the UNet1D model. The work demonstrates the feasibility of a robust and potentially applicable system in wearable devices for continuous sleep quality monitoring, including autonomic arousal events often missed in daily sleep quality analysis.
Il sonno rappresenta un processo fisiologico essenziale per la salute umana, la cui frammentazione si ripercuote sulle varie funzioni vitali che svolge per l’organismo e di conseguenza causa numerosi disturbi sistemici. Tra i principali meccanismi di disturbo vi è l’arousal autonomico, che comporta brevi e transitorie attivazioni del sistema nervoso autonomo non sempre accompagnate da segnali EEG e quindi difficili da rilevare. In questa tesi è stato sviluppato un framework automatico per il rilevamento non invasivo di questi eventi, basato esclusivamente sull’analisi del tachogramma (serie temporale di intervalli RR), con l’obiettivo futuro di un’applicazione in ambito wearable tramite segnali derivati dal PPG. Partendo dal dataset PhysioNet 2018, sono state implementate diverse fasi di pre-elaborazione, tra cui il filtraggio del segnale ECG, il rilevamento dei picchi R, la costruzione del tachogramma, la valutazione della qualità del segnale e la segmentazione in finestre di 60 secondi. Un elemento di particolare innovatività risiede nella costruzione del set di test: mentre i modelli sono stati addestrati su annotazioni EEG relative ad arousal a livello centrale, la fase di valutazione è stata condotta su un set indipendente annotato manualmente attraverso il solo segnale tachografico, per rilevare eventi di attivazione autonomica. Questo approccio simula realisticamente scenari d’uso senza EEG, come quelli legati ai dispositivi indossabili. Sono stati progettati e confrontati tre modelli di deep learning (CNN-LSTM, UNet1D e Transformer) per la classificazione binaria dei segmenti, valutati con soglie decisionali ottimizzate e utilizzando solo segmenti di alta qualità o tutti i segmenti indistintamente. Infine, una pipeline di post-elaborazione basata su filtri morfologici ha consentito una significativa riduzione dei falsi positivi (fino al 78%), con un aumento del punteggio F1 e dell’accuratezza, soprattutto per il modello UNet1D. Il lavoro dimostra la fattibilità di un sistema robusto e potenzialmente applicabile nei dispositivi indossabili per il monitoraggio continuo della qualità del sonno, compresi gli eventi di arousals autonomici che spesso sfuggono all’analisi quotidiana della qualità del sonno.
Deep learning-based detection of autonomic arousals from beat-to-beat intervals
Cantore, Andrea
2024/2025
Abstract
Sleep represents an essential physiological process for human health, the fragmentation of which impacts the various vital functions it performs for the body and consequently causes numerous systemic disorders. Among the main mechanisms of disruption is autonomic arousal, which involves brief and transient activations of the autonomic nervous system that are not always accompanied by EEG signals and are therefore difficult to detect. In this thesis, an automatic framework was developed for the noninvasive detection of these events, based solely on tachogram analysis (time series of RR intervals), with the future goal of application in the wearable domain via PPG-derived signals. Starting from the PhysioNet 2018 dataset, several pre-processing stages were implemented, including ECG signal filtering, R-peak detection, tachogram construction, signal quality assessment, and segmentation into 60-second windows. An element of particular innovativeness lies in the construction of the test set: while the models were trained on EEG annotations related to central arousal, the evaluation phase was conducted on an independent set manually annotated through tachographic signal alone, to detect autonomic activation events. This approach realistically simulates use scenarios without EEG, such as those related to wearable devices. Three deep learning models (CNN-LSTM, UNet1D, and Transformer) were designed and compared for binary classification of segments, evaluated with optimized decision thresholds, and on either using only high-quality segments or using all segments indiscriminately. Finally, a post-processing pipeline based on morphological filters allowed a significant reduction in false positives (up to 78%), with increased F1-score and accuracy, especially for the UNet1D model. The work demonstrates the feasibility of a robust and potentially applicable system in wearable devices for continuous sleep quality monitoring, including autonomic arousal events often missed in daily sleep quality analysis.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Cantore_Tesi_01.pdf
accessibile in internet per tutti a partire dal 25/06/2026
Descrizione: Testo Tesi
Dimensione
3.49 MB
Formato
Adobe PDF
|
3.49 MB | Adobe PDF | Visualizza/Apri |
|
2025_07_Cantore_Executive Summary_02.pdf
accessibile in internet per tutti a partire dal 25/06/2026
Descrizione: Executive Summary
Dimensione
851.75 kB
Formato
Adobe PDF
|
851.75 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240178