The aim of this Ph.D. thesis is to illustrate the research works conducted to design and develop advanced computational frameworks for analyzing electroencephalographic (EEG) signals to improve the early diagnosis of neurodegenerative diseases (NDs). Dementia is one of the leading causes of disability and death worldwide, and the detection of its initial phases remains a critical challenge both for prognostic and therapeutic purposes. The modern conceptualization of NDs, and particularly of Alzheimer’s disease, assumes cognitive decline to develop as a continuum, along which populations with still sufficient functional compensation could be targeted for early clinical trials. In this context, EEG signals can provide non-invasive and cost-effective biomarkers, holding potential for capturing neural dysfunctions associated with neurodegeneration. Nonetheless, the inherent complexity and variability of EEG result in significant challenges for accurate interpretation and analysis. This thesis addresses how Deep Learning (DL) models, particularly Transformers, and interpretability techniques can be leveraged for robust classification of EEG data, offering insights into subtle cognitive changes in preclinical and prodromal stages and overcoming the need for domain-specific expertise to extract consistent and reliable features. Furthermore, other approaches advancing the integration of computational neuroscience with Machine Learning (ML), including biophysical modeling of neural modulation in response to specific stimuli, are explored. In particular, the first part of the work presents a novel signal-based Deep Learning framework for distinguishing between subjective cognitive decline (SCD) and mild cognitive impairment (MCI) using resting-state EEG. The methods aim to capture prodromal signs of Alzheimer’s disease through a state-of-the-art Transformer model based on the mechanism of self-attention. To enhance clinical trustworthiness and translatability, the previously described method is then integrated with interpretability tools. Specifically, the role of self-attention within Transformer models is systematically explored to explain decision-making processes, providing greater transparency into the models’ focus on the input signals for differentiating SCD from MCI and proving that this information could be used to guide the identification of biomarkers of cognitive impairment in resting-state EEG. The second part of the research work presents computational methods for analyzing evoked responses, namely event-related potentials (ERP) and event-related (de)synchronization (ERD/ERS), in neurodegeneration, exploring motor resonance in early Parkinson’s disease, dynamic causal modeling for ERP classification, and the effects of sensory stimuli on electrophysiological responses in a Human-Robot Interaction scenario.
L'obiettivo della tesi di dottorato è quello di illustrare i lavori di ricerca svolti per la progettazione e lo sviluppo di framework computazionali avanzati per l'analisi dei segnali elettroencefalografici (EEG), al fine di migliorare la diagnosi precoce delle malattie neurodegenerative (ND). La demenza è una delle principali cause di disabilità e mortalità a livello globale, e l'identificazione delle sue fasi iniziali rimane una sfida critica sia per scopi prognostici che terapeutici. La moderna concettualizzazione delle ND, in particolare della malattia di Alzheimer, configura il declino cognitivo come un continuum, lungo il quale popolazioni con una compensazione funzionale ancora sufficiente potrebbero costituire target ideali per i trial clinici precoci. In questo contesto, i segnali EEG possono fornire biomarcatori non invasivi e a basso costo, con il potenziale di catturare disfunzioni neurali associate alla neurodegenerazione. Tuttavia, la complessità intrinseca e la variabilità dell’EEG comportano sfide significative per un'interpretazione e un'analisi accurate. La tesi affronta il modo in cui i modelli di Deep Learning (DL), in particolare i Transformers, e le tecniche di interpretabilità possano essere utilizzati per una classificazione robusta dei dati EEG, offrendo spunti sui cambiamenti cognitivi nelle fasi precliniche e prodromiche e superando la necessità di competenze specifiche per l'estrazione di caratteristiche consistenti e affidabili. Inoltre, vengono esplorati altri approcci che avanzano l'integrazione tra neuroscienze computazionali e Machine Learning (ML), includendo la modellazione biofisica della modulazione neurale in risposta a stimoli specifici. In particolare, la prima parte del lavoro presenta un nuovo framework di Deep Learning basato sui segnali, progettato per distinguere tra il declino cognitivo soggettivo (SCD) e l’impairment cognitivo lieve (MCI) utilizzando EEG resting-state. I metodi mirano a catturare i segni prodromici della malattia di Alzheimer attraverso un modello di Transformer basato sul meccanismo di self-attention. Per migliorare l'affidabilità e la traslabilità clinica, il framework descritto è stato integrato con strumenti di interpretabilità. Nello specifico, il ruolo del meccanismo di self-attention all'interno dei modelli Transformer è stato esplorato sistematicamente per spiegare i processi decisionali, fornendo maggiore trasparenza sul focus dei modelli sui segnali in ingresso per differenziare SCD da MCI e dimostrando che queste informazioni potrebbero essere utilizzate per guidare l'identificazione di biomarcatori di compromissione cognitiva nei segnali EEG a riposo. La seconda parte del lavoro di ricerca è focalizzata sull’implementazione di metodi computazionali per l'analisi delle risposte evocate, cioè dei potenziali evento-correlati (ERP) e della (de)sicronizzazione evento-correlata (ERD/ERS), nella neurodegenerazione. Vengono analizzati il meccanismo di risonanza motoria nelle fasi precoci della malattia di Parkinson, la modellazione causale dinamica per la classificazione degli ERP e gli effetti degli stimoli sensoriali sulle risposte elettrofisiologiche in uno scenario di interazione uomo-robot.
Advanced computational approaches for EEG-Based decoding of neurodegenerative diseases / Sibilano, Elena. - ELETTRONICO. - (2024).
Advanced computational approaches for EEG-Based decoding of neurodegenerative diseases
Sibilano, Elena
2024-01-01
Abstract
The aim of this Ph.D. thesis is to illustrate the research works conducted to design and develop advanced computational frameworks for analyzing electroencephalographic (EEG) signals to improve the early diagnosis of neurodegenerative diseases (NDs). Dementia is one of the leading causes of disability and death worldwide, and the detection of its initial phases remains a critical challenge both for prognostic and therapeutic purposes. The modern conceptualization of NDs, and particularly of Alzheimer’s disease, assumes cognitive decline to develop as a continuum, along which populations with still sufficient functional compensation could be targeted for early clinical trials. In this context, EEG signals can provide non-invasive and cost-effective biomarkers, holding potential for capturing neural dysfunctions associated with neurodegeneration. Nonetheless, the inherent complexity and variability of EEG result in significant challenges for accurate interpretation and analysis. This thesis addresses how Deep Learning (DL) models, particularly Transformers, and interpretability techniques can be leveraged for robust classification of EEG data, offering insights into subtle cognitive changes in preclinical and prodromal stages and overcoming the need for domain-specific expertise to extract consistent and reliable features. Furthermore, other approaches advancing the integration of computational neuroscience with Machine Learning (ML), including biophysical modeling of neural modulation in response to specific stimuli, are explored. In particular, the first part of the work presents a novel signal-based Deep Learning framework for distinguishing between subjective cognitive decline (SCD) and mild cognitive impairment (MCI) using resting-state EEG. The methods aim to capture prodromal signs of Alzheimer’s disease through a state-of-the-art Transformer model based on the mechanism of self-attention. To enhance clinical trustworthiness and translatability, the previously described method is then integrated with interpretability tools. Specifically, the role of self-attention within Transformer models is systematically explored to explain decision-making processes, providing greater transparency into the models’ focus on the input signals for differentiating SCD from MCI and proving that this information could be used to guide the identification of biomarkers of cognitive impairment in resting-state EEG. The second part of the research work presents computational methods for analyzing evoked responses, namely event-related potentials (ERP) and event-related (de)synchronization (ERD/ERS), in neurodegeneration, exploring motor resonance in early Parkinson’s disease, dynamic causal modeling for ERP classification, and the effects of sensory stimuli on electrophysiological responses in a Human-Robot Interaction scenario.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.