Major Depressive Disorder (MDD) is associated with significant impairments in executive functions, like inhibitory control. Despite growing interest in electroencephalography (EEG) for identifying with excellent temporal resolution the neural biomarkers of psychiatric conditions, few studies have exploited the innovative EEG microstates' framework to explore the functional brain dynamics during active cognitive tasks in MDD. This thesis investigates whether EEG microstate parameters derived from event-related potentials (ERPs) during a visuomotor Go/NoGo task can serve as reliable and interpretable biomarkers for distinguishing MDD patients from healthy controls. Forty participants (20 MDD patients, 20 healthy controls (HC)) performed a Go/NoGo task during the simultaneous acquisition of EEG and functional magnetic resonance imaging (fMRI). The task included two conditions using the right hand. In the excitatory condition, participants pressed a button whenever a square appeared, while in the inhibitory condition, they responded only to green squares and withheld responses to red ones. EEG preprocessing included artifact correction for MR-related and physiological noise, followed by ERP extraction that focused on N2 and P3 components. Statistical comparisons of ERPs were combined with EEG/ERP microstates and machine learning analyses. Grand-average ERPs were used to identify six microstate prototypes, which were then backfitted to each subject’s data to extract microstate metrics (e.g., global field power (GFP), duration, transition probabilities). These features were compiled into a structured dataset used to train multiple machine learning classifiers under a nested cross-validation framework. Among all the machine learning models available, seven supervised classifiers were chosen to evaluate their ability in distinguishing MDD patients from HC, based on EEG/ERP microstates features. To enhance interpretability, SHAP (SHapley Additive exPlanations) values were computed, allowing detailed insight into the contribution of each feature to model predictions. ERP analyses revealed significant group differences in frontal N2 and P3 components. In HC, the N2 and P3 showed stronger amplitudes and shorter latencies in the NoGo condition (non-target stimuli in inhibitory blocks) compared to Go/NoGo (target stimuli in inhibitory blocks), suggesting more efficient engagement of inhibitory mechanisms. In contrast, MDD patients exhibited reduced or absent N2 amplitude modulation and delayed P3 latencies, suggesting impaired conflict monitoring and slower cognitive processing. Microstate analysis identified six topographies from grand-average ERPs. HC displayed well-organized and diverse microstate sequences, especially during Go/NoGo trials, reflecting flexible neural transitions. In contrast, MDD participants showed more repetitive, less variable patterns, and reduced activation of key microstates such as MST2 (linked to conflict detection) and MST3 (linked to attentional control). Additionally, MST6 appeared exclusively in MDD and was not aligned with task-relevant ERP components, suggesting disease-specific neural correlates of task response. Among the classifiers tested, the MLP achieved the highest performance (accuracy = 0.748; AUC (Area under the curve) = 0.791). SHAP analysis revealed that the most influential features were the GFP of MST1 and MST2 and the transition probability from MST5 to MST6. Remarkably, lower GFP values in MST1 and MST2, particularly during NoGo trials, were strongly associated with MDD predictions, suggesting altered inhibitory processing. Overall, the results indicate that EEG/ERP microstate features, combined with interpretable machine learning techniques, may help uncover relevant neural alterations associated with MDD. This approach represents a promising step toward the development of interpretable EEG-based tools for clinical research. However, further validation in larger and independent samples is essential, along with the integration of multimodal data such as fMRI, to improve robustness and translational potential.
Il Disturbo Depressivo Maggiore (MDD) è associato a significative compromissioni delle funzioni esecutive, in particolare del controllo inibitorio. Nonostante il crescente interesse per l’elettroencefalografia (EEG) come strumento per identificare, con eccellente risoluzione temporale, biomarcatori neurali nelle condizioni psichiatriche, pochi studi hanno finora utilizzato l’innovativo approccio dei microstati per esplorare la dinamica funzionale del cervello durante compiti cognitivi attivi nei pazienti con MDD. Questa tesi si pone l'obiettivo di esplorare se i parametri dei microstati, derivati da potenziali evento-correlati (ERP) durante un compito visuomotorio di tipo Go/NoGo, possano rappresentare biomarcatori affidabili e interpretabili per distinguere i pazienti con MDD dai soggetti sani. Quaranta partecipanti (20 pazienti con MDD e 20 controlli sani, HC) hanno svolto un compito Go/NoGo durante la registrazione simultanea di EEG e risonanza magnetica funzionale (fMRI). Il task prevedeva due condizioni motorie eseguite con la mano destra: una eccitatoria, in cui i partecipanti dovevano premere un pulsante ogni volta che appariva un quadrato; e una inibitoria, in cui era richiesto di rispondere solo ai quadrati verdi e non premere il pulsante in presenza di quadrati rossi. L’analisi dell'EEG è iniziata con la correzione degli artefatti legati alla risonanza magnetica e al rumore fisiologico, seguita dall’estrazione degli ERP, con particolare attenzione alle componenti N2 e P3. Le analisi statistiche degli ERP sono state combinate con analisi di microstati EEG/ERP e tecniche di machine learning. A partire dagli ERP medi (grand average), sono stati identificati sei prototipi di microstato, che sono stati poi adattati ai dati individuali per estrarre metriche come la Global Field Power (GFP), la durata e le probabilità di transizione. Queste caratteristiche sono state raccolte in un dataset e utilizzate per addestrare diversi classificatori supervisionati tramite una strategia di validazione incrociata annidata (nested crossvalidation). Tra tutti i modelli di machine learning disponibili, sono stati selezionati sette classificatori supervisionati per valutare la capacità di distinguere i soggetti con MDD dagli HC, utilizzando le caratteristiche derivate dai microstati EEG/ERP. Per migliorare l’interpretabilità del modello, sono stati calcolati i valori SHAP (SHapley Additive exPlanations), che hanno permesso di analizzare nel dettaglio il contributo di ciascuna variabile. Le analisi ERP hanno evidenziato differenze significative tra i gruppi nelle componenti N2 e P3 frontali. Nei controlli sani, N2 e P3 mostravano ampiezze maggiori e latenze più brevi nella condizione NoGo (stimoli non target nei blocchi inibitori) rispetto alla Go/NoGo (stimoli target), suggerendo un coinvolgimento più efficiente dei meccanismi inibitori. Al contrario, nei pazienti con MDD si osservavano una modulazione ridotta o assente dell’N2 e latenze P3 più lunghe, indicando un’elaborazione cognitiva più lenta e un monitoraggio del conflitto compromesso. L’analisi dei microstati ha identificato sei topografie a partire dagli ERP medi. Gli HC mostravano sequenze di microstati organizzate e diversificate, in particolare nei trial Go/NoGo, riflettendo una maggiore flessibilità delle transizioni neurali. Al contrario, i soggetti MDD presentavano sequenze più ripetitive e meno variabili, con una ridotta attivazione di microstati chiave come MST2 (associato al rilevamento del conflitto) e MST3 (legato al controllo attentivo). Inoltre, MST6 compariva solo nei soggetti con MDD e non era associato a componenti ERP rilevanti per il compito, suggerendo la presenza di correlati neurali specifici della patologia. Tra i classificatori testati, il modello MLP ha ottenuto le prestazioni migliori (accuratezza = 0.748; AUC = 0.791). L’analisi SHAP ha indicato come caratteristiche più influenti la GFP di MST1 e MST2 e la probabilità di transizione da MST5 a MST6. Valori di GFP più bassi in MST1 e MST2, in particolare nei trial NoGo, risultavano fortemente associati alle predizioni di MDD, suggerendo un’alterazione nei processi inibitori. Nel complesso, i risultati mostrano che i microstati EEG/ERP, combinati con modelli di machine learning interpretabili, possono aiutare a individuare alterazioni neurali rilevanti associate al MDD. Questo approccio rappresenta un passo promettente verso lo sviluppo di strumenti EEG interpretabili per la ricerca clinica. Tuttavia, sono necessari ulteriori studi di validazione su campioni più ampi e indipendenti, oltre all’integrazione con dati multimodali come la fMRI, per migliorarne la robustezza e la trasferibilità clinica.
Uncovering the neural dynamics of inhibitory control in depression: a machine learning analysis of EEG/ERP microstates
Malosso, Micol
2024/2025
Abstract
Major Depressive Disorder (MDD) is associated with significant impairments in executive functions, like inhibitory control. Despite growing interest in electroencephalography (EEG) for identifying with excellent temporal resolution the neural biomarkers of psychiatric conditions, few studies have exploited the innovative EEG microstates' framework to explore the functional brain dynamics during active cognitive tasks in MDD. This thesis investigates whether EEG microstate parameters derived from event-related potentials (ERPs) during a visuomotor Go/NoGo task can serve as reliable and interpretable biomarkers for distinguishing MDD patients from healthy controls. Forty participants (20 MDD patients, 20 healthy controls (HC)) performed a Go/NoGo task during the simultaneous acquisition of EEG and functional magnetic resonance imaging (fMRI). The task included two conditions using the right hand. In the excitatory condition, participants pressed a button whenever a square appeared, while in the inhibitory condition, they responded only to green squares and withheld responses to red ones. EEG preprocessing included artifact correction for MR-related and physiological noise, followed by ERP extraction that focused on N2 and P3 components. Statistical comparisons of ERPs were combined with EEG/ERP microstates and machine learning analyses. Grand-average ERPs were used to identify six microstate prototypes, which were then backfitted to each subject’s data to extract microstate metrics (e.g., global field power (GFP), duration, transition probabilities). These features were compiled into a structured dataset used to train multiple machine learning classifiers under a nested cross-validation framework. Among all the machine learning models available, seven supervised classifiers were chosen to evaluate their ability in distinguishing MDD patients from HC, based on EEG/ERP microstates features. To enhance interpretability, SHAP (SHapley Additive exPlanations) values were computed, allowing detailed insight into the contribution of each feature to model predictions. ERP analyses revealed significant group differences in frontal N2 and P3 components. In HC, the N2 and P3 showed stronger amplitudes and shorter latencies in the NoGo condition (non-target stimuli in inhibitory blocks) compared to Go/NoGo (target stimuli in inhibitory blocks), suggesting more efficient engagement of inhibitory mechanisms. In contrast, MDD patients exhibited reduced or absent N2 amplitude modulation and delayed P3 latencies, suggesting impaired conflict monitoring and slower cognitive processing. Microstate analysis identified six topographies from grand-average ERPs. HC displayed well-organized and diverse microstate sequences, especially during Go/NoGo trials, reflecting flexible neural transitions. In contrast, MDD participants showed more repetitive, less variable patterns, and reduced activation of key microstates such as MST2 (linked to conflict detection) and MST3 (linked to attentional control). Additionally, MST6 appeared exclusively in MDD and was not aligned with task-relevant ERP components, suggesting disease-specific neural correlates of task response. Among the classifiers tested, the MLP achieved the highest performance (accuracy = 0.748; AUC (Area under the curve) = 0.791). SHAP analysis revealed that the most influential features were the GFP of MST1 and MST2 and the transition probability from MST5 to MST6. Remarkably, lower GFP values in MST1 and MST2, particularly during NoGo trials, were strongly associated with MDD predictions, suggesting altered inhibitory processing. Overall, the results indicate that EEG/ERP microstate features, combined with interpretable machine learning techniques, may help uncover relevant neural alterations associated with MDD. This approach represents a promising step toward the development of interpretable EEG-based tools for clinical research. However, further validation in larger and independent samples is essential, along with the integration of multimodal data such as fMRI, to improve robustness and translational potential.File | Dimensione | Formato | |
---|---|---|---|
2025_07_Malosso_Thesis_01.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
4.93 MB
Formato
Adobe PDF
|
4.93 MB | Adobe PDF | Visualizza/Apri |
2025_07_Malosso_Executive_Summary_02.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
972.51 kB
Formato
Adobe PDF
|
972.51 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/10589/239804