Deep Learning for RSVP-SSVEP Game

Advancements in machine learning, with the construct of deep learning, has made it possible to outperform many traditionally used algorithms. The healthcare system has seen a boom in machine learning research in areas such as classifying electroencephalography (EEG) signals for brain-computer interfaces (BCI). Two commonly used paradigms for BCI are the Rapid Serial Visual Presentation (RSVP) and Steady-State Visual Evoked Potentials (SSVEP). These paradigms can be invoked with visual stimuli and passively monitored with an EEG. While seeking to improve the performance of BCI systems and allow for more options for the user, recent studies have sought to utilize both RSVP and SSVEP. Current approaches to combine these two paradigms use a multimodal approach that trains at least two separate machine learning models to classify RSVP and SSVEP responses individually and then combines the results to create the final decision. This paper proposes a new deep learning algorithm using the unimodal architecture for RSVP-SSVEP BCI called RSNet. The SSVEP-RSVP BCI game used in this experiment is like the popular Bejeweled game developed by PopCap Games but was modified to elicit RSVP and SSVEP responses. This model was compared against multiple machine learning models; such as a bagged decision tree, fully connected deep neural network, EEGNet, and Compact-CNN to determine its performance relative to other architectures. RSNet was able to outperform all other models for the RSVP-SSVEP paradigm. RSNet expands the options for EEG classification and brings BCI systems one step closer to using the unimodal RSVP-SSVEP paradigm.