Theoretical and Computational Neuroscience
Author: Catalina María Galván | email: catalinamgalvan@gmail.com
Catalina M. Galván 1°, Rubén D. Spies 1°, Diego Milone 2°, Victoria Peterson 1°
1° Instituto de Matemática Aplicada del Litoral, IMAL, UNL, CONICET, Argentina
2° Instituto de Investigación en Señales, Sistemas e Inteligencia Computacional, sinc(i), FICH-UNL, CONICET, Argentina
Decoding algorithms for Brain-Computer Interfaces (BCIs) based on Electroencephalography (EEG) still experience several data-related limitations which bother their practical use. In this context, EEG simulation strategies play a crucial role in providing well-defined data for the development, validation and interpretation of EEG-BCI models. Moreover, in motor imagery (MI) BCIs for rehabilitation, large-scale subject-specific data is hardly accessible due to the high cost of human experiments and the reduced number of publicly available databases. This data volume has become more critical with the adveniment of deep learning (DL) models, whose performance strongly depends on how much training data is available. For such complex models, realistic simulation of MI-BCI recordings enables a strategy for data augmentation, boosting its performance and robustness in both a time- and cost-efficient fashion. In this work, we present a simulation framework, called PySimBCI, that can be used to generate realistic EEG-like signals based on specific model assumptions. We show that our artificially generated signals are electrophysiologically similar to real MI-EEG data. Experimental results exhibit that the performance of DL models can be effectively improved when the simulated data is introduced during the training process. Moreover, the proposed augmentation strategy yields a significant improvement over other state-of-the art augmentation methods.