PARADIGM-INDEPENDENT CLASSIFICATION ON MULTIDIMENSIONAL NEUROIMAGING DATASET USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Brain-Computer Interfaces (BCI) are gaining popularity with each day. A lot of research was carried out to improve both accuracy and usability of individual BCI paradigms, most popular in current literature being Motor Imagery(MI), Event-Related Potential (ERP)and Steady-State Visually Evoked Potential (SSVEP). However, individual BCI paradigms are limited in some areas such as number of classes in cases of MI and SSVEP paradigms, illiteracy in case of MI, or fatigue induction in users in case of ERP amd SSVEP paradigm. This study presents a new approach to designing BCI systems called paradigm-independent BCI, which solves these problems by allowing the subjects to use any of the three paradigms at any time. For each paradigm, the system processes EEG signals and feeds them through a deep learning model to get a representation vector, these vectors are then concatenated and fed to a final classifier that is used to decide which paradigm is currently being used on the fly. Average classification accuracy for three-class between-paradigm decoding was 97.51%(±3.39) in subject-independent training. Classification accuracy for seven-class within-paradigm decoding resulted in 87.9%(±4.51) accuracy. The results show that this framework works equally great in both subject-dependent and subject-independent contexts.