Objective. Deep learning algorithms employed in brain computer interfaces (BCIs) need large electroencephalographic (EEG) datasets to be trained. These datasets are usually unbalanced, particularly when error potential (ErrP) experiment are considered, being ErrP's epochs much rarer than non-ErrP ones. To face the issue of unbalance of rare epochs, this paper presents a novel, data balancing methods based on ARX-modelling. Approach. AutoRegressive with eXogenous input (ARX)-models are identified on the EEG data of the 'Monitoring error-related potentials' dataset of the BNCI Horizon 2020 and then employed to generate new synthetic data of the minority class of ErrP epochs. The balanced dataset is used to train a classifier of non-ErrP vs. ErrP epochs based on EEGNet. Main results. Compared to classical techniques (e.g. class weights, CW) for data balancing, the new method outperforms the others in terms of resulting accuracy (i.e. ARX 91.5% 88.3%
ARX-based EEG data balancing for error potential BCI
Farabbi, Andrea;Mainardi, Luca
2022-01-01
Abstract
Objective. Deep learning algorithms employed in brain computer interfaces (BCIs) need large electroencephalographic (EEG) datasets to be trained. These datasets are usually unbalanced, particularly when error potential (ErrP) experiment are considered, being ErrP's epochs much rarer than non-ErrP ones. To face the issue of unbalance of rare epochs, this paper presents a novel, data balancing methods based on ARX-modelling. Approach. AutoRegressive with eXogenous input (ARX)-models are identified on the EEG data of the 'Monitoring error-related potentials' dataset of the BNCI Horizon 2020 and then employed to generate new synthetic data of the minority class of ErrP epochs. The balanced dataset is used to train a classifier of non-ErrP vs. ErrP epochs based on EEGNet. Main results. Compared to classical techniques (e.g. class weights, CW) for data balancing, the new method outperforms the others in terms of resulting accuracy (i.e. ARX 91.5% 88.3%File | Dimensione | Formato | |
---|---|---|---|
Farabbi_2022_J._Neural_Eng._19_036023.pdf
Accesso riservato
:
Publisher’s version
Dimensione
4.39 MB
Formato
Adobe PDF
|
4.39 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.