Brain-computer interfaces (BCIs) offer an alternative communication channel between a subject and their environment: in BCIs based on Motor Imagery (MI), the subject’s ability to modulate their own EEG signal by imagining specific movements is exploited to control an external device. The present work focuses on a two-class MI-BCI experiment in which spatial covariance matrices were used as EEG signal descriptors, and classified by a model based on Riemannian differential geometry, the FgMDM classifier. This method extracts spatial information without the need for advanced spatial filtering, does not need feature selection, and allows for the manipulation of covariance matrices in their native space, i.e., in the manifold of symmetric positive-definite ones. Six healthy subjects participated in the experiment across four sessions to investigate whether a Riemannian approach to MI-BCIs can mitigate some of the limitations of traditional BCIs, namely, their lack of robustness and consequent lengthy calibrations. While one subject successfully used the same classifier throughout all training days, the models employed on the other participants had to be re-trained at least twice, possibly due to unstable MI activation patterns. For those subjects, the FgMDM classifier was not sufficient to overcome the drawbacks of state-of-the-art MI-BCIs, and additional adaptive techniques are needed. Nevertheless, the model explored herein proved able to extract information unavailable to traditional methods, confirming the value of geometry-aware approaches in decoding a user's intentions from EEG signals.
Le interfacce cervello-computer (BCI) offrono un canale di comunicazione alternativo per l'utente: nelle BCI basate sulla Motor Imagery (MI), la capacità del soggetto di modulare i propri segnali EEG, immaginando dei movimenti specifici, viene sfruttata per controllare un dispositivo esterno. Il presente lavoro si concentra su una MI-BCI a due classi, che categorizza le matrici di covarianza spaziale del segnale EEG grazie a un modello basato sulla geometria differenziale riemanniana, il classificatore FgMDM. Questo metodo estrae informazioni spaziali senza bisogno di tecniche avanzate, non necessita di selezione delle feature, e consente di maneggiare le matrici di covarianza nel loro spazio nativo, ovvero quello delle matrici simmetriche definite positive. Sei soggetti sani hanno partecipato all'esperimento in quattro sessioni ciascuno, per verificare se un approccio riemanniano alle MI-BCI possa mitigare alcuni dei limiti delle BCI tradizionali, ossia la loro mancanza di robustezza e le conseguenti lunghe calibrazioni. Un solo soggetto ha utilizzato con successo lo stesso classificatore in tutte le sedute. Per gli altri partecipanti è stato necessario aggiornare i modelli almeno due volte, probabilmente a causa dell'instabilità dei loro pattern di attivazione corticale; per questi soggetti il classificatore FgMDM non è stato sufficiente a superare le limitazioni dello stato dell'arte, e sarebbero necessarie tecniche adattive. Ciò nonostante, il modello considerato si è dimostrato in grado di estrarre informazioni non disponibili ai metodi tradizionali, confermando il valore degli approcci geometry-aware per l'interpretazione delle intenzioni di un utente a partire dai suoi segnali EEG.
Investigating Riemannian Geometry for Brain-Computer Interfaces based on Motor Imagery
FARCI, MATILDE
2024/2025
Abstract
Brain-computer interfaces (BCIs) offer an alternative communication channel between a subject and their environment: in BCIs based on Motor Imagery (MI), the subject’s ability to modulate their own EEG signal by imagining specific movements is exploited to control an external device. The present work focuses on a two-class MI-BCI experiment in which spatial covariance matrices were used as EEG signal descriptors, and classified by a model based on Riemannian differential geometry, the FgMDM classifier. This method extracts spatial information without the need for advanced spatial filtering, does not need feature selection, and allows for the manipulation of covariance matrices in their native space, i.e., in the manifold of symmetric positive-definite ones. Six healthy subjects participated in the experiment across four sessions to investigate whether a Riemannian approach to MI-BCIs can mitigate some of the limitations of traditional BCIs, namely, their lack of robustness and consequent lengthy calibrations. While one subject successfully used the same classifier throughout all training days, the models employed on the other participants had to be re-trained at least twice, possibly due to unstable MI activation patterns. For those subjects, the FgMDM classifier was not sufficient to overcome the drawbacks of state-of-the-art MI-BCIs, and additional adaptive techniques are needed. Nevertheless, the model explored herein proved able to extract information unavailable to traditional methods, confirming the value of geometry-aware approaches in decoding a user's intentions from EEG signals.| File | Dimensione | Formato | |
|---|---|---|---|
|
Farci_Matilde.pdf
Accesso riservato
Dimensione
14.09 MB
Formato
Adobe PDF
|
14.09 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/20.500.12608/87083