The P300 is one of the most widely used and studied Event-Related Potentials (ERPs) in Brain Computer Interfaces (BCIs). However, the existing multi-trial P300 detection methods suffer the drawback of being time-consuming and computationally complex. Therefore, in this thesis a single-trial detection strategy is explored through the analysis of the electroencephalographic (EEG) signals in the time-frequency domain and the implementation of machine learning algorithms. Employing an oddball paradigm, in which a series of low-probability target images are presented mixed with high-probability non-target ones, the P300 was elicited in 5 subjects between the ages of 22 and 27. The acquired signal was pre-processed through band-pass filtering, interpolation of the bad channels, Independent Component Analysis (ICA), re referencing, epoching, and baseline correction. Each trial was then associated with a feature vector, containing the corresponding Discrete Wavelet Transform (DWT) approximation coefficients in the 0-8 Hz frequency range from 7 channels located over the brain area associated to the P300. Feature selection was applied using Fisher Score and Kernel Principal Component Analysis (KPCA), to obtain 10 informative features for each trial. Finally, these features were used for the classification of the trials through a Support Vector Machine (SVM) classifier. For each subject, the strategy presented was tested 5 times using 5 different seeds for the random components of the algorithm. The performance of the classification was evaluated with the following statistical parameters: accuracy, sensitivity, specificity, and balanced accuracy. The respective average and standard deviation values of 77,5 ± 3,7 %, 76,7 ± 3,9 %, 77,6 ± 3,8 %, and 77,1 ± 3,5 % were obtained across the subjects. The time required to classify each trial was estimated to be lower than 250 ms. The results show that the algorithm was able to quickly assign a class to each trial based on the information from that trial alone. Moreover, the speed observed suggests that this strategy for the single-trial detection of the P300 is suitable for future real-time implementation in a BCI.
Il Potenziale Evento-Correlato (ERP) P300 è uno dei più utilizzati e studiati nelle Interfacce Cervello-Computer (BCI). I metodi che sfruttano l’informazione proveniente da molteplici trial per il rilevamento dell’ERP P300 sono dispendiosi in termini di tempo e complessi dal punto di vista computazionale. Pertanto, in questa tesi viene esplorata una strategia per il rilevamento del P300 sui singoli trial, attraverso l’analisi tempo-frequenza dei segnali elettroencefalografici (EEG) e l’uso di algoritmi di machine learning. La P300 è stata evocata utilizzando un paradigma “oddball”, in cui una serie di immagini target a bassa probabilità viene presentata in mezzo a immagini non-target ad alta probabilità. Il segnale acquisito in 5 soggetti (con età compresa tra i 22 e i 27 anni) è stato quindi pre-elaborato tramite filtraggio passa-banda, interpolazione dei canali difettosi, analisi delle componenti indipendenti (ICA), cambio di referenza, suddivisione delle epoche e baseline correction. Ogni trial è stato quindi associato a un vettore di feature, contenente i corrispondenti coefficienti di approssimazione della Trasformata Wavelet Discreta (DWT) nella banda di frequenza 0-8 Hz per 7 canali situati sull’area cerebrale associata alla P300. In seguito, è stata effettuata la selezione delle feature, utilizzando il Fisher Score e l’Analisi delle Componenti Principali con Kernel (KPCA), in modo da ottenere 10 feature per ogni trials. Queste feature sono state successivamente utilizzate per la classificazione dei trial, corrispondenti allo stimolo target e non-target, tramite l’algoritmo Support Vector Machine (SVM). Per ciascun soggetto, la strategia presentata è stata testata 5 volte utilizzando 5 seed diversi per le componenti randomiche dell’algoritmo. Le prestazioni della classificazione sono state stimate tramite i seguenti parametri statistici: accuratezza, sensibilità, specificità e accuratezza bilanciata. Fra i vari soggetti sono stati ottenuti i rispettivi valori medi di 77,5 ± 3,7 %, 76,7 ± 3,9 %, 77,6 ± 3,8 % e 77,1 ± 3,5 %. Il tempo necessario per classificare ciascuna prova è stato inferiore a 250 ms. I risultati dimostrano che l’algoritmo è stato in grado di assegnare rapidamente una classe a ciascun trial, basandosi esclusivamente sulle informazioni ricavate da quello stesso trial. Inoltre, la velocità osservata suggerisce che questa strategia per il rilevamento del P300 sui singoli trial sia adatta all’implementazione in tempo reale in una BCI.
A single-trial P300 detection strategy based on time-frequency analysis and machine learning for Brain Computer Interface applications
TONIUTTI, ALESSIA
2025/2026
Abstract
The P300 is one of the most widely used and studied Event-Related Potentials (ERPs) in Brain Computer Interfaces (BCIs). However, the existing multi-trial P300 detection methods suffer the drawback of being time-consuming and computationally complex. Therefore, in this thesis a single-trial detection strategy is explored through the analysis of the electroencephalographic (EEG) signals in the time-frequency domain and the implementation of machine learning algorithms. Employing an oddball paradigm, in which a series of low-probability target images are presented mixed with high-probability non-target ones, the P300 was elicited in 5 subjects between the ages of 22 and 27. The acquired signal was pre-processed through band-pass filtering, interpolation of the bad channels, Independent Component Analysis (ICA), re referencing, epoching, and baseline correction. Each trial was then associated with a feature vector, containing the corresponding Discrete Wavelet Transform (DWT) approximation coefficients in the 0-8 Hz frequency range from 7 channels located over the brain area associated to the P300. Feature selection was applied using Fisher Score and Kernel Principal Component Analysis (KPCA), to obtain 10 informative features for each trial. Finally, these features were used for the classification of the trials through a Support Vector Machine (SVM) classifier. For each subject, the strategy presented was tested 5 times using 5 different seeds for the random components of the algorithm. The performance of the classification was evaluated with the following statistical parameters: accuracy, sensitivity, specificity, and balanced accuracy. The respective average and standard deviation values of 77,5 ± 3,7 %, 76,7 ± 3,9 %, 77,6 ± 3,8 %, and 77,1 ± 3,5 % were obtained across the subjects. The time required to classify each trial was estimated to be lower than 250 ms. The results show that the algorithm was able to quickly assign a class to each trial based on the information from that trial alone. Moreover, the speed observed suggests that this strategy for the single-trial detection of the P300 is suitable for future real-time implementation in a BCI.| File | Dimensione | Formato | |
|---|---|---|---|
|
Toniutti_Alessia.pdf
accesso aperto
Dimensione
3.45 MB
Formato
Adobe PDF
|
3.45 MB | Adobe PDF | Visualizza/Apri |
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/20.500.12608/119610